com/abstract=2065904
1
The Role of Correlation Dynamics in Sector Allocation
Abstract
This paper assesses the role of correlation dynamics in meanvariance sector allocation. A
correlation timing framework is deployed to evaluate dynamic strategies against a static
constant covariance strategy and the nonparametric RiskMetrics covariance model. We find
using sector portfolios in three equity markets that correlation timing is rewarding. We
document timevariation, asymmetry and a structural break in sector correlations and show
that the predictability of conditional correlation models capturing such features is
economically relevant. The incremental value of correlation timing is more pronounced for
monthly rather than daily rebalancing and the gains of such strategies are not eroded by
transaction costs. Riskaverse monthly investors are willing to pay up to 1000 bp per annum
for switching from a static strategy to a dynamic correlation strategy. In contrast, the
RiskMetrics proves relatively trade intensive and sustains no net performance gains.
JEL Classification: C32, C52, C53, F21, G11, G15
Keywords: Dynamic sector allocation; Correlation timing; Portfolio performance; Utility
based evaluation; Transaction costs.
Electronic copy available at: http://ssrn.com/abstract=2065904
2
1. INTRODUCTION
Volatility and correlation among asset returns are central inputs to portfolio selection and
risk management. A burgeoning literature in financial economics has focused on time series
models for asset returns volatility and their co movement. The prolific research on the
relationship between stock returns and volatility has established that volatility is not only
time varying but also asymmetric, which implies that negative shocks have a greater impact
on future volatility than positive shocks of equal size. Black (1976), the first empirical study
on the riskreturn relation, found that future conditional stock volatility is negatively linked
to the current equity return and attributed this to the increased leverage surfacing when the
market value of a firm declines. Christie (1982) empirically corroborates the leverage
hypothesis through a positive relation between the firms debttoequity ratio and volatility.
On the other hand, Campbell and Hentschel (1992), the proponents of the volatility feedback
hypothesis, argue that negative unlike positive shocks increase volatility which has to be
compensated for by a sufficiently high expected return causing more volatility.
1
More recently, research focus has shifted to the estimation of the remaining of the
covariance structure of asset returns, and many different multivariate models, such as the
popular dynamic conditional correlation (DCC) model of Engle (2002) and the BEKK model
of Engle and Kroner (1995), have been proposed for this purpose. The literature on the
dynamics of asset correlations has by now formed the consensus that correlations change
through time (Bollerslev, Engle and Wooldridge, 1988). Erb, Harvey and Viskanta (1994)
link the time variation in correlation to cyclicality and show that correlations are higher in
recessions and lower when business cycles are out of phase. Correlation asymmetry is
another regularity that has been found in the second moment of equity returns, although
1
The leverage and volatility feedback interpretations of asymmetric volatility differ in regards to causalities: the
leverage hypothesis rests on the conjecture that return shocks lead to changes in conditional volatility, whereas
the volatility feedback theory contends that return shocks are caused by changes in conditional volatility. Bekaert
and Wu (2000) show that the two arguments are complementary and confirm that peaks in portfolio volatility
typically correspond to large market declines.
3
less attention has been paid to empirically capturing it. Longin and Solnik (2001) show that
correlations rise in bear markets, but not in bull markets. Ang and Bekaert (2002) document
the presence of a high volatilityhigh correlation regime in the US, UK and Germany, which
tends to coincide with a bear market. Cappiello, Engle and Sheppard (2006) find strong
evidence of asymmetry in conditional correlations of international equity and bond returns.
Although the economic rationale behind asymmetric effects is a relatively less researched
terrain, a few studies have put forth some potential explanations. Bekaert and Wu (2000)
attribute covariance asymmetry in financial assets to volatility feedback and show that
volatility increases dramatically upon large price declines but does not react to price
increases. Therefore, negative shocks among financial assets generate higher conditional
covariance than positive shocks. Bekaert, Harvey and Ng (2005) rationalize sudden jumps
in crossmarket correlations during crises to be a result of dependence on a common factor.
Structural breaks in correlations have also been documented. Billio and Pelizzon (2003)
find an increase in the level of conditional correlation of European equity markets in the
aftermath of the EMU and note that the effect had a fundamental impact on global markets.
Longin and Solnik (1995) and Goetzmann, Li and Rouwenhorst (2005) suggest that the level
and structure of global correlations shifted considerably over time, while Cappiello, Engle
and Sheppard (2006) find significant correlation increase postEMU not mirrored in
conditional volatility, which indicates greater market integration. Hyde, Bredin and
Nguyen (2007) document an Asian crisis structural break in the correlations between Asian
Pacific, European and US equity returns.
As timevariation, asymmetry and structural breaks are fast becoming stylized facts
of returns second joint moment (Cappiello, Engle and Sheppard, 2006) the evidence in
favour of multivariate conditional correlation models for characterizing co movement is
quite compelling. Continuous advances in correlation modeling could potentially facilitate
profitable investment or better risk management. But most extant studies on the evaluation
4
of conditional correlation estimators largely focus on statistical metrics. Engle and Sheppard
(2001) are the first to show that the DCC model outperforms the industry standard
RiskMetrics exponential smoother on the basis of residual normality and lower portfolio
standard deviations. Engle and Colacito (2006) show that the efficiency loss of mean
variance portfolios decreases as the estimated correlation approaches the true value and
further demonstrate that assuming constant correlation during volatile correlation phases is
costly, namely, as much as 40% of return can be dismissed, if the wrong conditional
correlation model were employed.
In this spirit, the question of whether it economically pays off to capture the stylized
facts of asset return correlations comes to the forefront. A natural way to assess the gains
from capturing covariance dynamics is by means of evaluating investment strategies
derived from them. Recent contributions have documented the economic value of
conditional volatility forecasts for asset allocation. Fleming, Kirby and Ostdiek (2001) find
that the predictability of conditional volatility models with nonparametric rolling
correlations is economically significant and robust to transaction costs. In a similar vein,
Della Corte, Sarno and Tsiakas (2009) document economic gains from the shorthorizon
predictive ability of economic fundamentals and forward premia on the volatility of
exchange rate returns. But less attention has been paid to the economic value of accounting
for correlation dynamics in equity investment. This is an important issue with direct
implications for investors and portfolio managers entertaining the use of conditional
correlation models. The typically low association between statistical accuracy and
profitability (Satchell and Timmermann, 1995) renders the question of whether there is
incentive for investors to opt for dynamic asset allocation based on conditional correlation
forecasts even more pertinent.
This is the first study to comprehensively assess the role of equity return correlation
dynamics in sector allocation. The contribution of this paper to the literature is twofold.
5
First, the benefits of correlation timing visvis the static constant covariance strategy are
assessed and the impact of transaction costs on the strategy performance is scrutinized.
Second, the role of correlation asymmetries and structural breaks in sector allocation is
statistically and economically evaluated. To this end, we consider alternative conditional
correlation models that allow for varying extents of dynamic co movement within a mean
variance framework designed to facilitate correlation timing. We adopt as nonparametric
benchmark the RiskMetrics exponential smoother that has been widely used in the industry
as a simple and viable way of estimating large dimensional covariance matrices. The
evaluation framework draws upon the seminal work of Fleming, Kirby and Ostdiek (2001),
where the relative economic value of dynamic strategies is gauged by their ability to
generate incremental utility to investors relative to static allocation. We assess the extent to
which dynamic strategies significantly improve portfolio return without unduly increasing
risk using the Sharpe Ratio. Finally, we investigate the effect of transaction costs and
rebalancing frequency on the performance of the correlation timing strategies.
The findings suggest that correlation timing is fruitful to sector investors. We find
economic value in capturing correlation changes including asymmetries and structural
breaks. Dynamic strategies provide superior riskadjusted returns and utilitybased
performance fees than the static constant covariance strategy and the volatility timing
strategy. Further, DCC models outperform the nonparametric RiskMetrics approach whose
value dissipates post transaction costs. The incremental gains of correlation timing are more
pronounced for monthly than daily rebalancing and are robust to reasonable transaction
costs. Monthly riskaverse investors are willing to pay a fee of up to 1000 bp to switch from
a static covariance strategy to a dynamic strategy triggered by correlation changes.
The remainder of the paper is organized as follows. Section 2 describes the data.
Section 3 delineates the conditional correlation models and the performance evaluation
framework. Section 4 discusses the empirical results. Finally, Section 5 concludes.
6
2. THE DATA
The empirical analysis is based on daily prices for ten sector indices from the Nikkei 225,
FTSEAll and S&P500 obtained from Thomson Reuters
(DataStream International
). The
ten sectors are: Energy (ENG), Basic Material (BML), Industrial (IND), Consumer Goods
(CGS), Health Care (HCR), Consumer Service (CSV), Telecommunication (TEL), Utility
(UTL), Financial (FIN) and Technology (TEC). The sample spans the period from July 1,
1996 to May 31, 2007, which amounts to a total of around 2700 local currency logarithmic
daily returns for each sector portfolio. The threemonth Japanese interbank loan rate, the
UK LIBOR, and the US Treasury bill rate are used as the risk free assets.
The descriptive statistics in Table 1 show positive mean daily returns for most sectors.
[Insert Table 1]
All daily returns are nonnormally distributed, particularly in the form of leptokurtosis. The
extent and direction of skewness differs across sectors and equity markets. Most of the
sector returns in the UK and US are significantly negatively skewed, whereas Japanese
sector returns are positively skewed.
2
The ADF test strongly rejects the hypothesis of a unit
root for all return series. The LjungBox Q statistic on daily and squared daily returns
portrays serial dependence and volatility persistence in virtually all sectors. The strong
evidence of volatility clustering supports the stylized fact that there is far more predictability
in conditional volatility than in return means.
The unconditional sector correlations over the sample period are significantly positive
in all markets. Appendix A1 reports the correlation matrices. The average unconditional
sector correlation level for Japan, UK and US is 54.9%, 44.3% and 52.5%, respectively.
3
Consumer services and industrials exhibit the highest correlations with other sectors at 61.6%
and 58.3%, respectively, while utilities are the least correlated with other sectors at 36.4%.
2
The lack of symmetry in the return distribution is consistent with previous findings (Bekaert and Wu, 2000).
3
The three mean correlations are significant with tstatistics 38.8, 35.3 and 27.26. The tstatistic is computed as
\(T2)/(1
2
) and follows a Studentt distribution with (T2) degrees of freedom.
7
Our empirical framework is designed to appraise the economic differences materializing
from rival correlation forecasting approaches. To this end, the sample is divided into an in
sample estimation period from July 1, 1996 to May 30, 2005 (T= 2274, 2266, 2209 days,
respectively, for the UK, US and JPN sector portfolios) and a holdout evaluation period from
June 1, 2005 to May 31, 2007 (T*= 484, 481, 472 days, respectively, for the UK, US and
Japanese sector portfolios) over which we generate onestepahead rolling covariance matrix
forecasts on the basis of a fixed window of lengthT.
3. METHODOLOGY
The analysis builds upon the recursive construction of optimal meanvariance sector
portfolios in the Japanese, UK and US markets and their outofsample performance
evaluation based on incremental utility and riskadjusted returns. For this purpose daily
sector correlation and volatility forecasts, the main inputs alongside expected returns for
dynamic asset allocation, are generated using the models outlined below.
3.1 The Conditional Covariance Structure
Let r
t
denote the day t logarithmic closetoclose return vector on n risky assets and
t1
be the
information set available at the end of day t1. The [n 1] conditional expected return
vector of r
t
is defined as
t
t,t1
= E[r
t
,
t1
] and H
t
H
t,t1
= E[(r
t

t
)
(r
t

t
),
t1
] is the
symmetric [n n] asset conditional covariance matrix. The return generating process is
conceptualized as r
t
= p
t
+E
t
12
e
t
, e
t
~N(u,1). At the open of day t, the investor formulates
a target return or target volatility quadratic optimization problem based on different
forecasts for the covariance matrix H
t
.
The basic estimator for H
t
is the insample unconditional covariance matrix E
0
= r
t
r
t
'
kept constant throughout the outofsample period. We make the covariance matrix
dynamic firstly by adopting in a multivariate setting the RiskMetrics estimator widely used
8
in the industry. The RiskMetrics or EWMA conditional covariance estimator is as follows
1 1 1
) 1 (
+ ' =
t t t t
H r r H (1)
, where the decay factor (0 < < 1) is = 0.94 for daily data and the insample covariance
matrix is taken as H
0
. The RiskMetrics model is computationally simple given its
nonparametric nature, however, its disadvantage is that it imposes identical covariance
dynamics to all assets as represented by the constant decay factor = 0.94. The conditional
variance estimates amount to those obtained by an IGARCH model.
Allowing for more flexibility while at the same time warranting parsimony we consider
conditional correlation models and extensions thereof that account for correlation
asymmetries and structural breaks. Conditional correlation models rely on decomposing
the conditional covariance into conditional standard deviations and conditional correlation.
The simplest model is the Constant Conditional Correlation (CCC) introduced by Bollerslev
(1990) which imposes time invariant correlation and covariance that changes over time
proportionally to the timevarying volatilities. The CCC model is estimated in two steps.
First, a univariate GARCH (p,q) model is fitted to each return series to generate the
conditional variance h
it
, i = 1,, n. Second, the conditional covariance is specified as
H
t
= D
t
R
D
t
(2)
, where ( )
nt t t
h h diag D ,...,
1
= and R is a positive definite [n n] correlation matrix
typically estimated by the unconditional insample correlation matrix.
The constant correlation assumption has been found to be too restrictive in several
empirical studies (Korner and Ng, 1998; Ang and Bekaert, 2002; Tse and Tsui, 2002), and so
the covariance decomposition in (2) has been extended to allow for dynamics in the
correlation matrix. Among the many specifications proposed for the evolution of R
t
the
Generalized Dynamic Conditional Correlation (GDCC) model of Engle (2002) is the most
popular. The GDCC model has the same first step as the CCC approach, but for each series
9
the standardized errors,
it
, are also generated alongside the conditional variance. In the
second step, the
it
are used to estimate the timevarying correlation matrix via
R
t
= (Q
t
*)
1
Q
t
(Q
t
*)
1
(3)
Q
t
= ( B Q B A Q A Q ' ' ) + A
t1
t1
A + BQ
t1
B
, whereQ = E[c
t
c
t
] is the unconditional covariance of standardized innovations, A and B are
[n n] diagonal parameter matrices, Q
t
* = diag(\q
it
,,\q
nt
) to ensure that R
t
has the structure
of a correlation matrix as long as the conditional covariance matrix Q
t
is positive definite.
4
The diagonal formulation in (3) poses assetspecific correlation dynamics but permits no
transmission of shocks between assets. The DCC model is obtained as a special case of the
GDCC where the parameter matrices are replaced by scalars, A= [a] and B= [b], and thus it
implies identical correlation dynamics among all assets.
The Asymmetric Generalized DCC (AGDCC) of Sheppard (2002), extends (3) by
allowing for asymmetries in the conditional covariance as follows
Q
t
= C + A
t1
t1
A + BQ
t1
B + Gq
t1
q
t1
G (4)
, where q
t
= I[
t
<0]
t
, indicates the elementbyelement Hadamard product, G is [n n]
diagonal parameter matrix, G N G B Q B A Q A Q C ' ' ' = andN = E[q
t
q
t
], where the
expectation is replaced by its sample analogue. Model (4) allows joint negative shocks to
have a stronger impact on future variances and correlations than positive shocks of the same
magnitude and nests the symmetric GDCC model. The Asymmetric DCC (ADCC) is the
restricted scalar version of (4), where A= [a], B= [b], C= [c] and G= [g]. Appendix A2 gives
technical details of the bivariate DCC and ADCC models and derives their loglikelihoods.
We also consider an extension of (4) which accommodates structural breaks in the long
run mean and in the dynamics of correlations (AGDCCBreak) as in Cappiello, Engle and
Sheppard (2006). The AGDCCBreak model accounts for two covariance regimes as follows
4
Q
t
will be positive definite with probability one if (Q AQA BQB) is positive definite.
10
Q
t
= d Q
t
1
+ (1d) Q
t
2
Q
t
j
= C
j
+ A
j
t1
t1
A
j
+ B
j
Q
t1
B
j
+ G
j
q
t1
q
t1
G
j
j =1, 2 (5)
where d is a break indicator defined as d = 1 for t < t, and 0 else. The corresponding scalar
model is the ADCCBreak, while the model collapses to GDCCBreak when G
j
= 0.
The News Impact Surface (NIS) for MGARCH, the analogue to the news impact curve
for univariate GARCH models, portrays how the conditional correlation of two assets reacts
to their joint past shocks (Kroner and Ng, 1998). The NIS correlation function (
1
,
2
) for the
AGDCC model takes the following approximate form under the assumption of linearity
) ( ) , (
j i j i
j
i j i ij j i
g g a a c f q q c c c c + + = (6)
, where c
ij
is the ij
th
element of the constant matrix C in (4). In the presence of asymmetry
parameter g
is significant and it is expected that joint bad news (
i
<
0,
j
<0) has a greater
impact on future correlation than joint good news, ceteris paribus.
Model estimation is by quasi maximum likelihood (QML). Inferences are based on
BollerslevWooldridge nonnormality robust standard errors (Bollerslev and Wooldridge,
1992). Individual significance tests are based on the associated tstatistics.
3.2 Dynamic Asset Allocation using Correlation Timing Strategies
A dynamic meanvariance framework is deployed to construct portfolios based on the
different daily covariance matrix forecasts. We consider an investor with a short or medium
term investment horizon who allocates funds across n risky assets plus a riskless security.
The investor solves the following constrained optimization problems at time t,
A. Maximize the conditional expected portfolio return subject to a target conditional
volatility
p
* (MaxR, hereafter).
_
max p
p
= w'
t
p
t
+(1 w'
t
I)r
]
s. t. o
p

= w'
t
E
t
w
t
(7)
B. Minimize the conditional portfolio variance subject to a target conditional expected return
11
p
* (MinV, hereafter).
_
min w'
t
E
t
w
t
s. t. p
p

= w'
t
p
t
+(1 w'
t
I)r
]
(8)
where
t
w is an [n 1] vector of portfolio weights on the risky assets, r
f
is the return on the
risk free asset, I is an [n 1] vector of 1s. In order to guarantee a feasible solution, no short
selling constraints are imposed. The optimal risky asset vector of weights is as follows
w
t
=
c
p

H
t
1
(
t

]
I)
_(
t

]
I)iH
t
1
(
t

]
I)
, for the Max R strategy,
w
t
=
(
p


]
I)H
t
1
(
t

]
I)
(
t

]
I)iH
t
1
(
t

]
I)
, for the MinV strategy,
, and the weight on the risk free asset is (1
t
w'
I).
When the conditional expected return
t
and conditional covariance H
t
are perceived
timevarying, investors will rebalance their portfolio weights following the dynamic
strategies outlined above. Emphasis is on the quality of the covariance forecast rather than
the correct modelling of expected return. Thus, the expected return is fixed (
t
= ) and
equal to the expost outofsample mean return, whereas the conditional covariance is
forecasted using the models in Section 3.1 to produce a sequence of optimal meanvariance
portfolios spanning the outofsample period. The CCC model amounts to a volatility
timing strategy and is adopted by investors who believe that changes in covariance are
driven by changes in volatility, while correlations are constant through time. The DCC
models can generate correlation timing strategies embedding various stylized facts of
correlations such as time variation, asymmetry and structural breaks. The static strategy is
based on the unconditional insample covariance matrix H
0
and is adopted by an investor
who believes that the conditional expected return and covariance are both constant, and
thus the optimal weights will be constant over time. The target return and variance
(p
p

, o
p

) are set at their expost mean values for an equally weighted portfolio of assets.
12
3.3 Performance Evaluation Framework
The adequacy of the dynamic strategies based on alternative covariance forecasts is judged
on the basis of their incremental utility visvis the benchmark static strategy. We follow
the utilitybased evaluation framework of Fleming, Kirby and Ostdiek (2001) drawing upon
the presumption that at a given point in time, one estimate of conditional covariance is
better than another if it leads to higher average utility. The quadratic utility function
introduced by West, Edison and Cho (1993) as a secondorder approximation to the
investors true utility in period t+1 is defined as
u
(w
t+1
) = w
t
R
p,t+1

uw
t
2
2
R
p,t+1
2
(9)
R
p,t+1
= (1 w
t
I)r
]
+w
t
r
t+1
, where W
t+1
is the wealth in period t+1, is the absolute risk aversion and R
p,t +1
is the
portfolio return at t+1. The expected endofperiod utility for a given level of initial wealth
W
0
and relative risk aversion
( ) / 1
t t t
W W o o = is estimated as
u
(w
t+1
) = w
0
[R
p,t+1

y
2(1+y)
R
p,t+1
2
(10)
We follow Fleming, Kirby and Osdtiek (2001) and assume constant relative risk aversion
levels of 1 and 10 to represent reasonably low and high riskaversions. The incremental
value of correlation timing visavis the static allocation is assessed by the return that would
render an investor indifferent between the two strategies as follows
j(R
d,t+1
) 
y
2(1+y)
(R
d,t+1
)
2
[ =
1

1
t=0
jR
s,t+1

y
2(1+y)
R
s,t+1
2
[
1

1
t=0
(11)
where R
d,t+1
and R
s,t+1
denote the returns for the dynamic and static strategies, respectively.
The equality in (11) implies that the investor would incur a daily expense A for the dynamic
strategy, and so we interpret A as the maximum performance fee (PF) in annualized basis
points the investor would be willing to pay to switch from the static to the dynamic strategy.
We statistically evaluate the riskadjusted performance of the strategies by assessing
13
the significance of the observed Sharpe Ratio (SR) differential of the dynamic strategy and
the static benchmark. In order to test the null hypothesis H
0
: (SR
d
 SR
s
) = 0 we employ the
asymptotic variance of the SR differential, Ior(SR
d]]
) = SR
d
 SR
s
, derived by Opdyke
(2007) under very general conditions as follows
Ior(SR
d]]
) = 1 +
SR
d
2
4
_
p
4d
o
d
4
 1_  SR
d
p
3d
o
d
3
+1 +
SR
s
2
4
_
p
4s
o
s
4
 1_ SR
s
p
3s
o
s
3
2 _p
ds
+
SR
d
SR
s
4
_
2d,2s
c
d
2
c
s
2
1] 
1
2
SR
d
1s,2d
c
s
c
d
2

1
2
SR
s
1d,2s
c
d
c
s
2
_ (12)
, where
nd,ms
=E[(R
d
E(R
d
)
n
(R
s
E(R
s
))
m
] is the joint central (n,m) moment of the joint
distribution of the two portfolio returns R
d
and R
s
. Unlike Lo (2002) where iid returns are
required, the asymptotic distribution in (12) requires only stationarity and ergodicity of
returns and is therefore valid under the more realistic conditions of timevarying volatilities,
serial correlation and noniid returns. A minimum variance unbiased estimator for this joint
moment is provided by the hstatistic of Rose and Smith (2002). Since the SR statistic is
asymptotically unbiased and normally distributed, the Central Limit Theorem implies that
I(SR
d]]
)
~
u
N(u, Ior(SR
d]]
)) (13)
Thus, the test statistic for equality of the SR of the competing strategies is z =
1(SR
d
SR
s
)
_vu(SR
di]]
)
.
3.4 Transaction Costs
Transaction costs play an important role when assessing the profitability of active trading
strategies. Accurate estimation of the size of transaction costs is challenging since it requires
information on the type of investor, the value of transaction, and the nature of the broker
(Della Corte, Sarno and Tsiakas, 2009). In order to sidestep these issues we follow Han (2006)
and compute breakeven transaction costs (BTC) per trade as the minimum proportional cost
that renders the investor with the quadratic utility function in (9) indifferent between the
dynamic strategy at hand and the static strategy. Assuming the investor pays a fixed
14
proportion (t ) of the value traded in each transaction, the total cost of a transaction in
period t can be represented as follows
_w
,t
w
,t1
1+
i,t
1+R
p,t
_
N
=1
(14)
, where w
,t1
1+
i,t
1+R
p,t
is the weight of asset i in the portfolio just before rebalancing in period t.
When assessing a dynamic strategy an investor who pays transaction costs lower than the
estimated BTC would favour the dynamic over the static strategy, and between two equally
performing dynamic strategies the one with higher BTC is preferable. BTC is proportional to
the value of each trade, and so it will be reported at the relevant trading frequency. Finally,
the turnover rate directly affects the posttransaction cost strategy performance to the extent
that realistic levels of transaction costs can negate any advantages associated with capturing
daily correlation fluctuations in asset allocation. We compute the average daily turnover
rate (TO) as the proportion of the portfolio value rebalanced each day, that is,
I0 = 1
I

, _w
,t
w
,t1
1+
i,t
1+R
p,t
_
N
=1
1

t=1
.
Sector index trading can be effectively replicated with Exchange Traded Funds (ETFs) at
a relatively low cost.
5
The total cost of investing in ETFs comprises the total expense ratio
(TER), the bidask spread, commission, and the cost of market impact. TER is the annual
management fee charged for operating expenses and is comparatively lower for ETFs than
mutual funds or other actively traded equity funds. The average level of TER for a US
investor trading SPDR sector ETFs listed on the NYSE is 20 bp. Sector ETFs traded on the
Tokyo Stock Exchange incur a TER of 2228 bp.
6
In the UK no sector ETFs have been
launched as yet, but the SPDR MSCI Europe sector ETFs with an associated 30 bp TER can
5
Since the introduction of SPDR ETF by State Street Global Advisor in 1993, which tracks the performance of the
S&P 500, the number of actively traded ETFs has ballooned due to their popularity among investors. In 2011Q4,
the number of traded ETFs was 1098 in the US and 1245 in Europe (Blackrock ETF Global Handbook). As trading
volume increased dramatically the cost of investing in ETFs has been brought down significantly in recent years.
6
The Select Sector SPDR Index Fund splits the S&P 500 into nine sectors traded individually on the NYSE.
Daiwa Asset Management provides 19 different sector ETFs for TOPIX companies (Daiwa ETFTOPIX).
15
be a reasonable proxy for UK sector indexlinked trading. For a daily trader the TER is
negligible (Jares and Lavin, 2004), however, for monthly traders it can play a role in the total
trading cost. Assuming investors will pay a fraction of the TER proportional to the length of
the holding period a monthly sector ETF trader should bear a TER of up to 2 bp per trade.
For frequent traders of ETFs the trading cost depends primarily on the bidask spread.
Previous studies on ETFs typically use the close bidask prices to calculate the bidask
spread (Engle and Sarkar, 2002; Jares and Lavin, 2004). However, bidask spreads tend to be
wider at the end of the trading day than during the trading hours since traders face a higher
risk that their order might not be executed (Foucault, 1999; McInish and Wood, 1992). The
higher bidask spread of the last trade(s) can also be attributed to the introduction of the
closing auction on most of the exchanges.
7
This implies that using the endofday bidask
spread would inflate the actual trading cost. To circumvent this issue we use intraday price
quotes and compute the bidask spread on day t as BidAsk
t
= min(P
jt
)/LowP
t
for j =
1,,M intraday intervals, where min(P
jt
) is the smallest intraday price change observed
during day t and LowP
t
is the lowest price level on day t. The estimated average BidAsk
t
for SPDR US Sector ETFs ranges from 1.8 to 4.5 bp, with the exception of financials for which
it is slightly higher. The Daiwa JPN TOPIX Sector ETFs average BidAsk
t
is found to be 28
bp, whereas for the SPDR MSCI Europe Sector ETFs it is typically around 48 bp.
8
Similar to trading other securities, investors who buy ETFs need to pay commissions to
their brokers. However, since the commission fee is charged on the number of trades, the
percentage cost of commission can be very small when the volume of transactions per trade
is very high. Dellva (2001) shows that the commission for SPDR ETFs is only $20 per trade,
7
Admati and Pfleiderer (1988) argue that the bidask spread or terms of trade is determined by the number of
informed traders in the market. If the number of informed traders increases, the terms of trade will worsen
especially when the opinions of these informed traders are diversified. The closing auction will attract more
informed traders into the price discovery process as they possess more information about the underlying asset
and could form a better strategy in the auction, therefore, the terms of trade will worsen.
8
The average bidask spread at the last trade is found to be considerably larger than during the day, for instance,
it is at 45 bp for the SPDR US Industrial ETF and 54 bp for the DAIWA JPN ETFTOPIX Construction & Material.
16
which means the cost of commission is only 2 bp if one invests $100,000 at once. For
institutional or large individual investors the percentage cost of commission can be even
lower. Furthermore, in order to reduce the trading cost and attract more investors, some of
the ETF providers have in recent years introduced commissionfree ETFs.
9
Finally, transaction costs are affected by the cost of market impact, that is, the influence
of an investors decision on the market price of the underlying asset. The cost of market
impact tends to be lower for highly liquid assets, which implies the behaviour of one trader
does not have a big influence on the price. Given the high trading volume of popular ETFs,
the market impact of a single institutional investor is almost negligible.
10
Jares and Lavin
(2004) do not account for the cost of market impact when considering ETF trading costs,
however, practitioners still count it as one of the cost components. We take a conservative
approach and use a cost of market impact of 2 bp as suggested for large cap index ETFs.
11
In summary, the total trading cost of ETFs for a daily trader can be represented by the
sum of bidask spread and the cost of market impact as both the commission and TER are
negligible. In the context of sectorlinked ETFs this amounts to approximately 7 bp per day
for US investors, 30 bp per day for Japanese investors and 50 bp per day for European sector
investors. For monthly traders we follow the industry practice to incur a TER proportional
to the holding period and assume an additional of cost of 2 bp. The total trading cost in each
case will be used as a benchmark to assess the practical feasibility of the dynamic strategies.
4. EMPIRICAL RESULTS
4.1 The Dynamics of Sector Correlations
We start by examining the estimation results of the different multivariate conditional
9
For instance, Fidelity and Vanguard Asset Management have waived the commission for some of their widely
traded ETFs.
10
In the last quarter of 2011, the average daily trading volume of SPDR US Industrial and SPDR US Financial was
30.15 and 120.5 million, respectively.
11
See Frontier Investment Management report at http://www.frontierim.com/uploads/frontierinvestmentmanagement 
whenisaternotater.pdf.
17
correlation estimators. The covariance matrices for the three domestic sector portfolios are
estimated over the entire sample period, July 1, 1996 to May 31, 2007. The firststep in the
estimation of conditional correlation models is to specify the univariate conditional variance.
Thus, for each daily return series we fit a GARCH (1,1) model. To account for asymmetry in
the impact of news on conditional volatility we also fit Nelson (1991)s EGARCH (1,1,1). We
find no evidence of volatility asymmetry and the GARCH specification is favoured for all
sector index returns by the Akaike (AIC) and Schwarz (SIC) information criteria.
12
Given the
conditional volatilities, the conditional correlations are estimated using equations (2) to (5)
for each sector portfolio. Following the pertinent literature (Baele, 2005; Billio and Pelizzon,
2003 inter alios), we introduce a structural break at the onset of the European Monetary
Union (EMU) on January 1, 1999, when all the EMU members irrevocably fix their exchange
rate and the Euro is introduced to replace the national currency. The radical transform of
the European money market influenced the interdependence of the EMU member
economies and that of the closely integrated UK, US and Japanese markets.
13
Figure 1 shows the rolling daily unconditional correlations between financialindustrial
and energyutility sectors over the sample period.
[Insert Figure 1]
The graphs confirm the extensive time variation in sector correlations and provide some
support for a regime change after 1999. Correlations tend to bottom out in 1999 and
gradually recover over the next few years. The low correlations observed in 1999 coincide
with the upward trend in the three markets, while the subsequent correlation rise relates to
the bear market observed globally over 20002002. This is in line with previous studies,
12
The results of the univariate conditional volatility models are available from the authors upon request.
13
The transformation of the European money market also affected the US dollar cash flows, through its impact
on Eurodollar interest rates. The change in the term structure influenced the value of the dollar and
consequently impacted the US economy. The fundamental changes in the European and US money markets
affected the domestic equity markets and further transmitted to Japan due to the strong degree of globalization
(Hamao, Masulis and Ng, 1990; Koutmos and Booth, 1995). Nonetheless, the time taken for the effects of the
money market transformation in Europe to transmit to the US and then further influence Japan is less clearcut.
18
which support the conjecture that correlations among financial assets tend to increase
during bear markets and decrease during bull markets (see, inter alios, Longin and Solnik,
1995; Ang and Bekaert, 2002). Default risk plays an important role in explaining the increase
in sector correlations during bear markets. As a common risk factor to all sectors, default
risk tends to be higher in economic downturn which increases the sector exposure to this
common risk factor, and therefore results in higher sector correlations. Dllmann, Scheicher
and Schmieder (2007) find a positive link between asset correlations and borrowing size,
which tends to increase during recessions.
Empirical likelihood ratio (LR) tests reported in Table 2 give strong evidence for the
existence of a structural break in sector correlation dynamics in each of the three markets.
[Insert Table 2]
Moreover, the diagonal GDCC appears more appropriate than its scalar counterpart for all
sector portfolios the LR tests reject the null hypothesis of identical correlation dynamics
across assets. Asymmetry in sector correlations is also borne out by a significant increase in
the value of the loglikelihood function upon inclusion of the asymmetric term. On the other
hand, the AIC and SIC information criteria that tradeoff fit and parsimony point towards
the scalar DCC, the most parsimonious among the dynamic correlation models.
The parameter estimates for the conditional correlation models are set out in Table 3.
Most parameters are statistically significant at the conventional levels.
[Insert Table 3]
We find evidence of correlation asymmetry in the majority of sector correlations indicated
by the significance of the asymmetry parameter g. The findings also indicate a change in the
dynamic structure of conditional correlations following the introduction of the EMU. The
significance of the difference in the parameters that dictate the correlation evolution in the
19
two periods is assessed by the Welchs Students ttest.
14
The results provide evidence that
the dynamics of sector correlations changes significantly postEMU. The longrun effect of
joint news on correlations (b) seems to rise, while the impact of recent news on the shortrun
correlation dynamics (a) drops significantly. The degree of persistence in conditional
correlation, measured by a
+ b + g in the scalar models and a
2
+ b
2
+ g
2
in the diagonal models,
also undergoes a structural break. Conditional correlations become more persistent after the
introduction of the EMU. For instance, the persistence in conditional correlation for the US
sectors is 0.920 (DCCBreak) and 0.935 (ADCCBreak) in the prebreak period, and rises to
0.998 and 0.997, respectively, in the postbreak period. Asymmetric effects in correlations
tend to dampen significantly postEMU, particularly in the US sectors.
Time series graphs of the conditional correlations obtained from the AGDCCBreak
model are presented in Figure 2 for the financialindustrial and utilityenergy sector pairs.
[Insert Figure 2]
In all three markets sector correlations decrease sharply at the end of 1998 and recover to
their preEMU level by the end of 2002. To gain more insight into the impact of the
structural break and asymmetries on correlation dynamics, we plot the correlation news
impact surface (NIS) derived from the AGDCCBreak model. Figure 3 shows the NIS for
financials  industrials in the three markets during the pre and post1999 periods.
[Insert Figure 3]
The NIS corroborates the evidence obtained from the parameter estimates in Table 3
regarding the presence of asymmetric effects and structural break in sector correlations.
There is notable asymmetry in correlations in the pre1999 period, which implies that
14
The test statistic is t =
X
1
X
2

S
12
, where X
1
and X
2
are the corresponding coefficients for the pre and postbreak
period, S
12
= _
S
1
2
N
1

S
2
2
N
2
is the pooled variance of the (X
1
X
2
) differential, S
1
and S
2
are the variances of
parameters X
1
and X
2
, N
1
and N
2
are the sample sizes. The degreeoffreedom of the Studentt distribution is
_
S
1
2
N
1
+
S
2
2
N
2
]
2
_
S
1
2
N
1
_
2
N
1
1
+
_
S
2
2
N
2
_
2
N
2
1
. The results of the test are available from the authors upon request.
20
correlation increases dramatically when both sectors move in the negative direction, while
the impact of joint positive news is relatively lower. However, asymmetry in correlation
dynamics is largely eliminated post1999. The results suggest that allowing for asymmetric
effect and structural break in correlation is important, particularly for the US and UK sectors.
4.2 Timing the Correlation Signals
We now turn to investigate whether capturing the temporal evolution of correlations along
with asymmetric effects and breaks is economically significant. To this end, we assess the
performance of the dynamic strategies based on alternative covariance matrix estimators
against static sector allocation. The strategies are contrasted within the economic framework
outlined in Section 3.3 that gauges their ability to generate excess riskadjusted returns and
incremental utility. The meanvariance sector portfolios are recursively constructed based
on the onestepahead covariance matrix forecasts obtained from equations (1) to (5), while
the static portfolio is based on the insample unconditional covariance matrix.
Table 4 presents the outofsample evaluation of the correlation timing strategies against
the static constant covariance benchmark under the MaxR scheme.
[Insert Table 4]
First we appraise the standard portfolio performance measures. Reported for each sector
portfolio is the annualized mean portfolio return (), return standard deviation (), Sharpe
Ratio (SR) and SR differential (SR) relative to the static strategy, and the associated p
values for Opdykes (2007) test of equality of SRs. A significant test statistic indicates
rejection of H
0
: SR
d
=SR
s
in favour of the alternative that the dynamic strategy yields a higher
SR than the static. Bold denotes the best performing model under a given criterion.
The results suggest that the dynamic sector allocation strategies are generally able to
generate higher rewardtorisk ratios than the static constant covariance strategy. In the UK,
any conditional correlation strategy entails significantly better SR than the static strategy.
21
The best model for the UK sector portfolio turns out to be the DCC which accrues significant
incremental gains in riskadjusted return of SR = 1.340 in excess of the static strategy, while
for the US the CCC model that maintains constant correlation achieves the highest
significant increase in the SR of 0.698 relative to the static allocation. In Japan dynamic
allocation increases the SR albeit not significantly so.
We now turn attention to the economic value of the covariance forecasting models on
the basis of annualized performance fee (PF) of the strategy at hand visvis the static
benchmark. We find large and positive performance fees across all portfolios providing
overwhelming evidence that the dynamic strategies outperform the static constant
covariance strategy in all three markets. Among the various strategies, the EWMA entails
the largest gains for sector investors with PFs ranging from 1282 bp to 1892 bp depending on
the portfolio and riskaversion coefficient. That is, a highly risk averse Japanese sector
investor would be willing to pay up to a maximum of 1892 bp per annum to reap the
benefits of the dynamic EWMA covariance strategy; a similar US investor would be willing
to pay up to 1584 bp. Focusing on conditional correlation models, there is evidence that
accounting for correlation asymmetries and possibly breaks enhances performance fees. The
top performing models are AGDCCBreak for the US and Japan, and AGDCC for the UK.
Table 5 reports the performance evaluation of the dynamic versus the static portfolios
under the MinV portfolio construction scheme.
[Insert Table 5]
The results suggest that the dynamic strategies outperform the static benchmark on the basis
of lower portfolio volatility and significantly higher riskadjusted returns. There is strong
evidence that dynamic strategies significantly increase the SR of UK and US sector portfolios.
From the viewpoint of riskadjusted returns the best performer for the UK is the DCC and
for US it is the CCC, both with significant SR differentials of SR = 1.591 and SR = 1.007.
In Japan dynamic allocation entails SR gains, but these are not statistically significant.
22
In terms of utilitybased performance fees, the dynamic strategies still outperform
the static strategy for the UK and US sector portfolios. A highly riskaverse UK investor is
willing to pay up to 136 bp per annum to switch from the static constant covariance strategy
to the topranked DCC correlation timing strategy. In the US, such an investor is willing to
pay up to 259 bp per annum to switch from the static strategy to one where rebalancing is
driven by CCC volatility forecasts. In contrast, the Japanese sector investor does not seem to
benefit from either correlation or volatility timing consistent with the evidence derived from
SRs. In the case of Japan, the outperformance of the static strategy is mainly driven by the
relatively higher return. Dynamic strategies do manage to achieve a lower volatility than
the static strategy (e.g. = 9.58% ADCC vs. = 10.91% Static), however, this is outweighed
by the relatively lower portfolio return ( = 20.46% ADCC vs. = 21.91% Static).
4.3 Turnover Rate and BreakEven Transaction Costs
The empirical results obtained thus far suggest that the dynamic strategies outperform the
static constant covariance strategy in terms of rewardtorisk ratio and performance fees for
riskaverse investors with a quadratic utility function. However, active trading strategies
are prone to high turnovers and their performance can therefore be substantially impeded
by transaction costs. The monthly turnover volume (TO) for each strategy can be seen in
Tables 4 and 5. The TO of the static strategy that only rebalances in order to maintain
constant weights is 0.20 0.71 (MaxR), or equivalently 20%  71% of total portfolio value,
and 0.13 0.37 (MinV). The monthly turnover for the conditional correlation timing
strategies is considerable, ranging at 4.75 14.55 (MaxR) and 2.08 7.74 (MinV) across
models and portfolios. In case the EWMA approach is used for forecasting the covariance
matrix, the TO increases even further to 17.24 25 (MaxR) and 4.69 10.01 (MinV). The
strategy with the lowest TO employs the CCC model that rebalances in response to volatility
changes but does not react to changes in correlation.
23
The differences in turnover rate among dynamic strategies have important
implications for their posttransaction cost economic value, which is summarized by the
breakeven transaction cost (BTC). The results in Tables 4 and 5 indicate that highly risk
averse US sector investor using the trade intensive EWMA model faces a BTC of 6.17 bp
(1.31 bp) per trade under MaxR (MinV), which is lower than the transaction cost of 7 bp
paid for trading US Sector ETFs. On the other hand, a US investor using the low turnover
CCC model faces a higher and economically plausible BTC of 17.34 bp (9.26 bp) per trade
under MaxR (MinV). For DCCtype models the BTCs are still higher than the actual level
of transaction costs. Therefore, US portfolio managers opting for conditional correlation
models rather than the EWMA model can reap net performance gains due to the lower
number of trades of the former. Nonetheless, in the UK and Japan the BTC of conditional
correlation models with daily rebalancing is below the indicated average trading costs for
sectorlinked ETFs. Despite the positive PFs, the relatively high transaction costs in these
two markets cast doubt on the practical feasibility of sector correlation timing.
4.4 The Impact of Rebalancing Frequency on the Performance of Dynamic Strategies
As seen in the previous section, day traders engaging in dynamic correlation strategies face
small BTCs, which imply that the incremental gains of the dynamic strategies relative to the
static strategy dissipate for realistic levels of transaction costs. Lower rebalancing frequency
can reduce the turnover, which is negatively related to the revision interval, and potentially
allow investors to effectively implement the dynamic strategies.
In order to investigate the impact of rebalancing frequency on the performance of the
dynamic asset allocation strategies, we repeat the analysis for monthly and weekly horizons.
To this end, we rebalance the portfolio daily based on the daily covariance matrix forecast
and hold the new portfolio for a mday holding period, where m = 21 for a monthly horizon
and m = 5 for a weekly horizon. This overlapping approach assumes that, on each trading
24
day, the investor will hold multiple portfolios simultaneously, each formed one day apart
but only one of the m portfolios will be revised.
15
The overall dayt return is calculated as
the equally weighted average return of the m portfolios held on day t, and the weights are
kept at 1/m throughout the outofsample period. The turnover ratio of the total asset
holding on each day is equal to the turnover ratio of the revised portfolio multiplied by its
weight. The advantage of the overlapping approach is twofold. First, it enables the use of
all onedayahead covariance forecasts. Second, it eliminates the bias arising from the day of
the week effect, and accounts for performance variability from the choice of rebalancing day.
Tables 6 to 9 set out the impact of lowering the rebalancing frequency on the outof
sample performance of the dynamic strategies. Tables 6  7 present the results of weekly
rebalancing and Tables 8  9 those of monthly rebalancing.
[Insert Tables 6  9]
Our findings suggest that reducing the rebalancing frequency enhances the riskadjusted
performance of both static and dynamic portfolios. In line with Fleming, Kirby and Ostdiek
(2003) and De Pooter, Martens and van Dijk (2008), the observed higher SRs are driven by
considerable increases in the mean portfolio return, while the increases in portfolio volatility
are slight.
16
The incremental gains of dynamic strategies relative to the static benchmark are
more pronounced at lower rebalancing frequencies. Monthly volatility/correlation timing
generates PFs ranging from 690 bp to 2281 bp (MaxR, Table 8) a notable increase from 554
bp to 1892 bp obtained in the case of daily rebalancing. This can be attributed to the fact that
dynamic portfolios derive more benefit from longer revision intervals than static portfolios,
which implies that investors are prepared to pay higher PFs to switch from static to dynamic
15
The overlapping method to evaluate the performance of stock picking techniques with different rebalancing
frequencies is inspired by Jegadeesh and Titmans (1993) and Rouwenhorsts (1998) early work on portfolio
trading strategies. A nonoverlapping approach, which assumes the investor only holds one portfolio and
rebalances it at the end of each holding period, is used in Fleming, Kirby and Ostdiek (2003).
16
De Pooter, Martens and van Dijk (2008) assess the impact of rebalancing frequency on the performance of daily
and intradaybased dynamic covariance portfolios and find that lower frequencies yield higher performance fees
and emphasize more the benefits of intraday data for covariance matrix estimation.
25
strategies when rebalancing less often. For instance, the SR of the Japanese sector portfolio
based on the CCC model increases from 2.37 with daily rebalancing to 2.45 with monthly
rebalancing, but the benefits for the corresponding static portfolio are meagre.
Furthermore, the decrease in turnover volume rate when switching from daily to
weekly or monthly rebalancing is quite dramatic for both dynamic and static strategies. The
TO of weekly portfolios is less than half the TO of the daily portfolios, while revising the
portfolios on a monthly basis reduces the TO even further to around a quarter of the TO of
their daily counterparts. As an example, the daily dynamic strategy based on CCC forecasts
under MaxR has a TO rate of 8.68, 5.12 and 4.75, respectively for Japanese, UK and US
sector portfolios, whereas the TO rate of the corresponding monthly portfolios is curtailed to
1.69, 1.08 and 0.93.
A direct implication of the enhanced performance fees and the lower turnover is the
higher BTCs associated with the lower rebalancing frequencies, which imply that dynamic
portfolios are more likely to maintain posttransaction cost benefits over the static
benchmark if they are revised less frequently. Akin to daily rebalancing, weekly rebalancing
renders correlation timing feasible only in the US market. However, as borne out in Table 8
MaxR monthly rebalancing generates BTCs markedly above realistic levels of transaction
costs pointing towards the use of dynamic strategies in all markets. Depending on the
model and riskaversion, the BTCs generated by portfolios based on conditional correlation
models range from 36 bp to 92 bp per trade. Monthly MinV correlation timing (Table 9) is
feasible for US investors with BTCs of about 25 bp. Although the BTCs of such strategies are
not convincingly above the rather high realistic levels of transaction costs in the UK and
Japan, the improvement entailed by less frequent rebalancing is also noticeable in these
markets. In contrast, the EWMA unanimously fails to beat the static strategy (negative PFs).
In order to more directly evaluate the effect of rebalancing frequency on the dynamic
strategies we compute for any given model the maximum return an investor is willing to
26
forfeit to switch from daily to weekly and to monthly rebalancing. Table 10 presents these
performance fees for each strategy, model and level of investor riskaversion.
[Insert Table 10]
The results suggest that weekly and monthly rebalanced portfolios based on dynamic
correlations outperform their daily counterparts by generating positive PFs regardless of the
riskaversion level. This finding is in line with De Pooter, Martens and Dijk (2008) in the
context of intraday covariance estimates. For the MaxR strategy we find that the PF for
switching from daily to weekly rebalancing is between 56 bp and 642 bp for sector investors
using dynamic correlation strategies, while these performance fees increase further to the
range of 70 bp to 795 bp when switching from daily to monthly rebalancing. Similar
inferences are derived for the MinV strategy, with the exception of the US where the results
are mixed. US correlation timing based on the AGDCC model is favoured by less frequent
rebalancing, but this is not true for other models.
The results suggest that dynamic strategies based on conditional correlation models
outperform those derived from the industry standard EWMA when transaction costs are
considered. Second, a longer holding period makes the gains of the dynamic strategies
relative to the static benchmark more pronounced. Lower rebalancing frequency also
increases the BTC per trade of the dynamic strategies, thus making them economically
plausible. Liketolike strategy comparison at different rebalancing frequencies suggests
that correlation timing adds more economic value when implemented monthly than daily.
4.5 Robustness Tests
It can be argued that dynamic strategies outperform the static strategy because the chosen
target return setting is unfavourable to the latter. We test the robustness of the relative
performance of the dynamic correlation strategies visvis the static strategy by examining
the sensitivity to different target return levels. To this end, we compare the riskreturn
27
tradeoff of the dynamic portfolios against the capital market line (CML) derived from the
static benchmark. Portfolios on the CML provide the highest possible SR among all efficient
static portfolios, and so if the dynamic portfolios in Tables 4  5 outperform the static CML,
then their superior performance is robust to the choice of target return.
We construct the static CML by generating static MinV portfolios using the expost
mean return vector and the insample unconditional covariance matrix, where the highest
target return is the maximum return achieved by the daily dynamic strategies, while the
lowest target return is zero.
17
Figure 4 illustrates the riskreturn performance of the daily
dynamic portfolios in Tables 4 and 5 against the CML of the static covariance strategy.
[Insert Figure 4]
The results suggest that portfolios based on dynamic volatility and correlation forecasts
outperform the static CML by providing higher SRs in all three markets. Therefore, the
results validate the robustness of the performance gains associated with the correlation
timing strategies for the whole spectrum of efficient portfolios produced by any given model.
Finally, we assess the asset allocation implication of the covariance matrix forecast
accuracy for the entire efficient frontier of the dynamic portfolios over the outofsample
period. Engle and Colacito (2006) show that portfolios based on more accurate variance
covariance forecasts provide lower volatility for a given level of expected return, that is, they
are more efficient. Therefore, the forecast accuracy of a correlation forecasting model can be
assessed by comparing the resulting efficient frontier to that of the static strategy and to the
expost frontier which assumes perfect foresight of volatilities and correlations.
In order to carry out this check, we track the efficient frontier of each strategy using the
onedayahead covariance matrix forecasts and a set of target returns. For each target return,
the actual (realized) daily portfolio volatility is computed, based on the optimal weights
17
The expost mean return vector is generated by averaging the daily return vector of sector indices over the out
ofsample period, while the expost covariance matrix is the outofsample average of the daily realized
covariance matrices, H
t
= r
t
r
t
.
28
derived from the covariance matrix forecast, the expost return vector and expost
covariance matrix. The target return is constant over the outofsample period, while the
associated portfolio volatility changes according to the weighting scheme conduced from the
covariance matrix. We average the realized daily portfolio volatilities for each target return
over the outofsample period and them into an aggregate efficient frontier. The efficient
frontier derived from the outofsample unconditional covariance matrix is also produced as
the optimal frontier the best strategy can ever achieve expost assuming perfect foresight.
The estimated efficient frontiers of the static and the dynamic strategies are presented in
Figure 5, alongside the expost optimal frontier.
[Insert Figure 5]
In all three markets the efficient frontiers of the dynamic strategies embrace the static
frontier, which corroborates that all efficient dynamic portfolios can achieve a better risk
return tradeoff than the efficient static portfolios. Another finding that emerges robustly
from this analysis is that the dynamic efficient frontier is closer to the expost optimal
frontier than the static frontier is. This suggests that the daily conditional covariance matrix
forecasts are more accurate than the unconditional constant covariance benchmark for all
efficient portfolios. The results, therefore, confirm that the outperformance of the dynamic
strategies is attributed to more accurate covariance matrix forecasts and is not an artefact of
the portfolio construction strategy at hand.
5. CONCLUSIONS
Forecasting the asset return correlation is crucial for portfolio selection and risk
management and thus, various estimation approaches have been developed for this purpose.
In this study, we investigate the economic value of correlation timing in sector allocation.
We benchmark the dynamic strategies against a static constant covariance strategy within a
utilitybased framework that accounts for the impact of transaction costs. To this end, we
29
gauge the relative merits of various multivariate conditional correlation forecasts by looking
at the riskreturn profile and incremental utility of the optimal meanvariance portfolios.
The empirical results suggest that the predictability of conditional correlation models is
economically relevant. The findings based on sector portfolios in three markets suggest that
the dynamic asset allocation strategies outperform the static strategy in terms of risk
adjusted returns and further accrue economic value. Although daily correlation timing has
been found to afford gains to investors over and above the static constant covariance
benchmark, transaction costs have been shown to impede the strategy performance.
Correlation timing conducted monthly can, however, outperform static allocation and
generates economically plausible breakeven transaction costs. Our findings further suggest
that the incremental gains of dynamic strategies relative to the static strategy are more
pronounced when considering monthly holding periods instead of daily and that exploiting
correlation dynamics is more beneficial for longer horizon investors. Conditional correlation
models outperform the nonparametric RiskMetrics EWMA approach whose value added is
negated by transaction costs. Overall, we show that the stylized facts of asset correlations
such as timevariation, and in some cases asymmetry and structural breaks are mirrored by
substantial economic payoffs in the context of dynamic sector allocation.
30
REFERENCES
Admati, A.R. and Pfleiderer, P. (1988) A theory of intraday patterns: Volume and price
variability. Review of Financial Studies 1:340.
Ang, A. and Bekaert, G. (2002) International asset allocation with regime shifts. Review of
Financial Studies 15:11371187.
Baele, L. (2005) Volatility spillover effects in European equity markets. Journal of Financial
and Quantitative Analysis 40:373401.
Bekaert, G. and Wu, G. (2000) Asymmetric volatility and risk in equity markets. Review of
Financial Studies 13:142.
Bekaert, G., Harvey, C. and Ng, A. (2005) Market integration and contagion. Journal of
Business 78:3969.
Billio, M. and Pelizzon, L. (2003) Volatility and shocks spillover before and after EMU in
European stock markets. Journal of Multinational Financial Management 13:323340.
Black, F. (1976) Studies of stock price volatility changes. Proceedings of the 1976 meetings of the
American Statistical Association, Business and Economics Statistics Section, 177181.
Bollerslev, T. (1990) Modeling the coherence in shortrun nominal exchange rate: a
multivariate Generalized ARCH approach. Review of Economics and Statistic 72:49850.
Bollerslev, T., Engle, R. and Wooldridge, J. (1988) A capital asset pricing model with time
varying covariances. The Journal of Political Economy 96:116131.
Bollerslev, T. and Wooldridge, J. (1992) Quasimaximum likelihood estimation and inference
in dynamic models with timevarying covariance. Economic Review 11:143172.
Campbell, J.Y. and Hentschel,L. (1992) No news is good news: An asymmetric model of
changing volatility in stock returns. Journal of Financial Economics 31:281318.
Cappiello, L., Engle, R. and Sheppard, K. (2006) Asymmetric Dynamics in the correlations of
global equity and bond returns. Journal of Financial Econometrics 4:537572.
Christie, A. (1982) The stochastic behaviour of common stock variances: Value, leverage, and
interest rate effects. Journal of Financial Economics 10:407432.
De Pooter, M., Martens, M. and van Dijk, D. (2008) Predicting the daily covariance matrix for
S&P 100 stocks using intraday data but which frequency to use? Econometric Review 27:199
229.
Della Corte, P., Sarno, L. and Tsiakas, I. (2009) An economic evaluation of empirical
exchange rate models. Review of Financial Studies 22:34913530.
Dellva, W.L. (2001) Exchangetraded funds not for everyone. Journal of Financial Planning
14:110122.
31
Dllmann, K., Scheicher, M. and Schmieder, C. (2007) Asset correlations and credit portfolio risk
An empirical analysis. Discussion Paper Series 2: Banking and Financial Studies No 13/2007.
Deutsche Bundesbank.
Engle, R. (2002) Dynamic conditional correlation: a simple class of multivariate generalized
autoregressive conditional heteroskedasticity models. Journal of Business and Economic
Statistics 20:339350.
Engle, R. and Colacito, R. (2006) Testing and valuing dynamic correlations for asset
allocation. Journal of Business and Economic Statistics 24:238253.
Engle, R. and Kroner, K. (1995) Multivariate simultaneous GARCH. Econometric Theory
11:122150.
Engle, R. and Sarkar, D. (2002) Pricing exchange traded funds. NYU Working Paper No.SDRP
0211.
Engle, R. and Sheppard, K. (2001) Theoretical and empirical properties of Dynamic Conditional
Correlation MVGARCH. Working paper No. 200115, University of California, San Diego.
Erb, C., Harvey, C. and Viskanta, T. (1994) Forecasting international equity correlations.
Financial Analysts Journal 50:3245.
Fleming, J., Kirby, C., and Ostdiek, B. (2001) The economic value of volatility timing. Journal
of Finance 56:329352.
Fleming, J., Kirby, C., and Ostdiek, B. (2003) The economic value of volatility timing using
realized volatility. Journal of Financial Economics 67:473509.
Foucault, T. (1999) Order flow composition and trading costs in a dynamic limit order
market. Journal of Financial Markets 2:99134.
Goetzmann, W., Li, L. and Rouwenhorst, K. (2005) Longterm global market correlations.
Journal of Business 78:138.
Hamao, Y. Masulis, R. and Ng, V. (1990) Correlations in price changes and volatility across
international stock markets. Review of Financial Studies 3:281307.
Han, Y. (2006) Asset allocation with a high dimensional latent factor stochastic volatility
model. Review of Financial Studies 19:237271.
Hyde, S., Bredin, D. and Nguyen, N. (2007) Correlation dynamic between AsiaPacific, EU
and US stock returns. International Finance Review 8:3961.
Jares, T.E. and Lavin, A.M. (2004) Japan and Hong Kong exchangetraded funds (ETFs):
Discounts, returns, and trading strategies. Journal of Financial Services Research 25:5769.
Jegadeesh, M. and Titman, S. (1993) Returns to buying winners and selling losers:
Implications for stock market efficiency. Journal of Finance 48:6591.
Koutmos, G. and Booth, G.G. (1995) Asymmetric volatility transmission in international
32
stock markets. Journal of International Money and Finance 14:747762.
Kroner, K. and Ng, V. (1998) Modeling asymmetric comovements of asset returns. Review of
Financial Studies 11:817844.
Lo, A.W. (2002) The statistics of Sharpe ratios. Financial Analysts Journal 58:3652.
Longin, F. and Solnik, B. (1995) Is the correlation in international equity returns constant:
19601990? Journal of International Money and Finance 14:326.
Longin, F. and Solnik, B. (2001) Extreme correlation of international equity markets. Journal
of Finance 56:649676.
McInish, T.H. and Wood, R.A. (1992) An analysis of intraday patterns in bid/ask spreads for
NYSE stocks. Journal of Finance 47:753764.
Nelson, D. (1991) Conditional heteroskedasticity in asset returns: a new approach.
Econometrica 59:34970.
Opdyke, J.D. (2007) Comparing Sharpe ratio: So where are the pvalues. Journal of Asset
Management 8:308336.
Rose, C. and Smith, M. (2002) Mathematical Statistics with Mathematica, SpringerVerlag,
New York.
Rouwenhorst, K.G. (1998) International momentum strategies. Journal of Finance 53:267284.
Satchell, S., and Timmermann, A. (1995). An assessment of the economic value of nonlinear
foreign exchange rate forecasts. Journal of Forecasting 14:477497.
Sheppard, K. (2002) Understanding the dynamics of equity covariance. Manuscript, UCSD
Tse, Y. and Tsui, A. (2002) A multivariate generalized autoregressive conditional
heteroskedasticity model with timevarying correlations. Journal of Business and Economic
Statistics 20:351362.
Welch, B. L. (1947) The generalization of students problem when several different
population variances are involved. Biometrika 34:2835.
West, K. D., Edison, H. J. and Cho, D. (1993) A utility based comparison of some models of
exchange rate volatility. Journal of International Economics 35:2345.
33
Appendix A1  Unconditional Daily Sector Correlations
Appendix A2 The DCC and ADCC MGARCH models for two assets
The Dynamic Conditional Correlation (DCC) MGARCH model, for a twoasset case, follows
the process
r
t
=H
t
1/2
c
t
, c
t
~ N(0,1)
1, 1, 2,
1, 2, 2,
t t t t
t
t t t t
h h h
H
h h h
(
(
=
(
where the conditional variances are specified as
h
1,t
= e
1
+ o
1
r
2
1,t1
+ 
1
h
1,t1
h
2,t
= e
2
+ o
2
r
2
2,t1
+ 
2
h
2,t1
Sector Indices
ENG BML IND CGS HCR CSV TEL UTL FIN TEC
Japan
ENG 1
BML 0.61 1
IND 0.44 0.78 1
CGS 0.38 0.65 0.80 1
HCR 0.46 0.66 0.59 0.58 1
CSV 0.54 0.82 0.79 0.72 0.71 1
TEL 0.29 0.51 0.65 0.57 0.44 0.61 1
UTL 0.35 0.42 0.30 0.31 0.48 0.46 0.22 1
FIN 0.50 0.77 0.70 0.61 0.56 0.76 0.54 0.32 1
TEC 0.27 0.58 0.83 0.68 0.42 0.64 0.67 0.14 0.58 1
UK
ENG 1
BML 0.48 1
IND 0.41 0.59 1
CGS 0.33 0.45 0.49 1
HCR 0.43 0.39 0.37 0.35 1
CSV 0.44 0.60 0.58 0.48 0.49 1
TEL 0.35 0.39 0.41 0.30 0.39 0.64 1
UTL 0.38 0.37 0.33 0.28 0.46 0.47 0.35 1
FIN 0.52 0.57 0.56 0.48 0.60 0.72 0.57 0.49 1
TEC 0.26 0.39 0.41 0.30 0.24 0.60 0.51 0.22 0.47 1
US
ENG 1
BML 0.47 1
IND 0.41 0.72 1
CGS 0.35 0.69 0.72 1
HCR 0.40 0.52 0.63 0.53 1
CSV 0.35 0.65 0.79 0.73 0.59 1
TEL 0.32 0.47 0.59 0.52 0.51 0.63 1
UTL 0.45 0.39 0.42 0.37 0.42 0.37 0.37 1
FIN 0.37 0.64 0.77 0.69 0.63 0.76 0.60 0.45 1
TEC 0.23 0.43 0.65 0.55 0.39 0.68 0.55 0.24 0.58 1
34
and
t
= q
12,t
/(q
1,t
q
2,t
)
1/2
is derived from
q
1,t
=q
1
(1  a  b) + a c
2
1,t1
+ b q
1,t1
q
2,t
=q
2
(1  a  b) + a c
2
2,t1
+ b q
2,t1
q
12,t
=q
12
(1  a  b) + a c
2
1,t1
c
2
2,t1
+ b q
12,t1
whereq
1
,
q
2
, andq
12
equal the unconditional variance (q
1
andq
2
) and covariance (q
12
) of
the two assets returns, respectively.
The Asymmetric Dynamic Conditional Correlation (ADCC) MGARCH model gives
higher conditional correlation when both assets experience negative shocks compared to
positive shocks. The ADCC estimator has the following structure in a two assets case
r
t
=H
t
1/2
c
t
, c
t
~ N(0,1)
1, 1, 2,
1, 2, 2,
t t t t
t
t t t t
h h h
H
h h h
(
(
=
(
where the conditional variances are specified as
h
1,t
= e
1
+ o
1
r
2
1,t1
+ 
1
h
1,t1
h
2,t
= e
2
+ o
2
r
2
2,t1
+ 
2
h
2,t1
and
t
= q
12,t
/(q
1,t
q
2,t
)
1/2
is derived from
q
1,t
=q
1
(1  a b g/2) + a c
2
1,t1
+ b q
1,t1
+ g d
1,t1
c
2
1,t1
q
2,t
=q
2
(1  a b g/2) + a c
2
2,t1
+ b q
2,t1
+ g d
2,t1
c
2
2,t1
q
12,t
=q
12
(1  a  b) +q
3
g + a c
2
1,t1
c
2
2,t1
+ b q
12,t1
+ g (d
1,t1
c
1,t1
) (d
2,t1
c
2,t1
)
d
i,t
= I(c
i,t
< 0)
where d
1,t
and d
2,t
are dummies for r
1,t
and r
2,t
that assume value 1 whenever these variables
are negative and 0 otherwise, and the coefficient g relies on the assumption that
1
and
2
have a symmetric distribution. q
12
andq
3
are the unconditional covariance of the assets
returns and the their asymmetric components
(d
1,t1
c
1,t1
) (d
2,t1
c
2,t1
).
QuasiMaximum Likelihood (QML) Estimation
The twostep estimation of DCC and ADCC MVGARCH model assume that
r
t

t1
~ N(0, H
t
) ~ N(0, D
t
R
t
D
t
)
The normality assumption of r
t
gives rise to a loglikelihood function. Without the
normality assumption, the estimator will still have the QML interpretation. The
loglikelihood for this estimator can be written as
35
L =
1
1
2
T
t =
(n log(2t) + log H
t
 + log R
t
 + r
t
1
t
H
r
t
)
=
1
1
2
T
t =
(n log(2t) + log D
t
R
t
D
t
 + r
t
1 1 1
t t t
D R D
r
t
)
Since the standardized residual,
t
= u
t
/ \ h
t
=
1
t
D
r
t
, the loglikelihood function can be
expressed as
L =
1
1
2
T
t =
(n log(2t) + 2 log D
t
 + log R
t
 +
t
1
t
R
t
)
=
1
1
2
T
t =
(n log(2t) + 2 log D
t
 + r
t
1 1
t t
D D
r
t

t
t
+ log R
t
 +
t
1
t
R
t
)
It is clear that there are two separate parts of the loglikelihood function, the volatility
part containing D
t
and the correlation part containing R
t
. This gives rise to the two stage
estimation procedure. In the first stage, each of D
t
can be considered as an univariate
GARCH model, therefore the loglikelihood of the volatility term is simply the sum of the
loglikelihoods of the individual GARCH equations for the involved return series
L =
1
1
2
T
t =
(n log(2t) + 2 log D
t
 + r
t
1 1
t t
D D
r
t
)
=
1
1
2
T
t =
(n log(2t) + 2 log D
t
 + r
t
2
t
D
r
t
)
=
1
1
2
T
t =
(n log(2t) +
1
n
i =
(log(h
it
) +
2
it
r / h
it
) )
=
1
1
2
n
t =
(T log(2t) +
1
T
i =
(log(h
it
) +
2
it
r / h
it
) )
In the second stage, the parameters of the correlation evolution are estimated using the
specified loglikelihood of the correlation part, conditioning on the parameters estimated in
the first stage likelihood
L
C
=
1
1
2
T
t =
( log R
t
 +
t
1
t
R
t

t
t
)
The properties of the QML estimation and related test statistics in dynamic models that
jointly parameterize conditional means and conditional covariances are discussed in
Bollerslev and Wooldridge (1992). It should be noted that the two step estimation of the
likelihood function means that estimation is inefficient, though consistent (Engle and
Sheppard, 2001; Engle, 2002).
36
Figure 1: Rolling Unconditional Correlations
Time series graphs of daily unconditional correlation for financial (FIN) industrial (IND) sectors and utility
(UTL) energy (ENG) sectors estimated using a onemonth rolling window.
Panel A. Japanese sectors
Panel B. UK sectors
Panel C. US sectors
0.60
0.40
0.20
0.00
0.20
0.40
0.60
0.80
1.00
07/1996 07/1997 07/1998 07/1999 07/2000 07/2001 07/2002 07/2003 07/2004 07/2005 07/2006
FININD UTLENG
0.60
0.40
0.20
0.00
0.20
0.40
0.60
0.80
1.00
07/1996 07/1997 07/1998 07/1999 07/2000 07/2001 07/2002 07/2003 07/2004 07/2005 07/2006
FININD UTLENG
0.60
0.40
0.20
0.00
0.20
0.40
0.60
0.80
1.00
07/1996 07/1997 07/1998 07/1999 07/2000 07/2001 07/2002 07/2003 07/2004 07/2005 07/2006
FININD UTLENG
37
Figure 2: Conditional Correlation Dynamics
Time series graphs of daily conditional correlation for financial (FIN) industrial (IND) and utility (UTL)
energy (ENG) sectors obtained from the AGDCCBreak model with a structural break on January 1, 1999.
Panel A. Japanese sectors
Panel B. UK sectors
Panel C. US sectors
0.00
0.10
0.20
0.30
0.40
0.50
0.60
0.70
0.80
0.90
1.00
07/1996 07/1997 07/1998 07/1999 07/2000 07/2001 07/2002 07/2003 07/2004 07/2005 07/2006
FININD UTLENG
0.00
0.10
0.20
0.30
0.40
0.50
0.60
0.70
0.80
07/1996 07/1997 07/1998 07/1999 07/2000 07/2001 07/2002 07/2003 07/2004 07/2005 07/2006
FININD UTLENG
0.00
0.10
0.20
0.30
0.40
0.50
0.60
0.70
0.80
0.90
1.00
07/1996 07/1997 07/1998 07/1999 07/2000 07/2001 07/2002 07/2003 07/2004 07/2005 07/2006
FININD UTLENG
38
Figure 3: Correlation News Impact Surfacess
Conditional correlation News Impact Surface for financial (FIN) industrial (IND) sectors derived from the
AGDCCBreak with a structural break on January 1, 1999.
Panel A. News Impact Surface for the Japanese FININD sectors
PreEMU period PostEMU period
Panel B. News Impact Surface for the UK FININD sectors
PreEMU period PostEMU period
Panel C. News Impact Surface for the US FININD sectors
PreEMU period PostEMU period
39
Figure 4: Dynamic Strategies and the Static Capital Market Line
The graphs demonstrate the relative performance of the dynamic covariance strategies against the capital market
line (CML) derived from the static constant covariance strategy. The annualized riskreturn tradeoff of the
conditional correlation MaxR and MinV portfolios in Tables 4 and 5 are plotted as black rectangles. The grey
rectangles represent the dynamic portfolios based on the EWMA covariance model.
Panel A. Japanese sector portfolios
Panel B. UK sector portfolios
Panel C. US sector portfolios
0 5 10 15 20 25 30
0
10
20
30
40
50
60
Portfolio Volatility (%)
P
o
r
t
f
o
l
i
o
R
e
t
u
r
n
(
%
)
EWMA
Conditional Correlation
Models
0 5 10 15 20 25
0
10
20
30
40
50
Portfolio Volatility (%)
P
o
r
t
f
o
l
i
o
R
e
t
u
r
n
(
%
)
Conditional Correlation
Models
EWMA
0 5 10 15 20 25
0
5
10
15
20
25
30
35
Portfolio Volatility (%)
P
o
r
t
f
o
l
i
o
R
e
t
u
r
n
(
%
)
Conditional Correlation
Models
EWMA
Static CML
Static CML
Static CML
40
Figure 5: Efficient frontiers of the Dynamic and Static Covariance Strategies
The graphs illustrate the efficient frontier of the dynamic strategies based on the EWMA covariance model and
the AGDCCBreak conditional correlation model aggregated over the outofsample period. In each graph, the
aggregated efficient frontier of the given dynamic portfolio (solid line) is plotted together with the static (dotted
line) covariance portfolio and the expost (dashed line) optimal efficient frontier.
Panel A. Japanese sector portfolios
EWMA AGDCCBreak
Panel B. UK sector portfolios
EWMA AGDCCBreak
Panel C. US sector portfolios
EWMA AGDCCBreak
13 14 15 16 17 18 19 20 21 22 23
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
Portfolio Volatility (%)
P
o
r
t
f
o
l
i
o
R
e
t
u
r
n
(
%
)
EWMA
Static
Ex.Post
13 14 15 16 17 18 19 20 21 22 23
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
Portfolio Volatility (%)
P
o
r
t
f
o
l
i
o
R
e
t
u
r
n
(
%
)
AGDCCBreak
Static
Ex.Post
8 10 12 14 16 18 20 22 24 26 28
0
5
10
15
20
25
30
35
Portfolio Volatility (%)
P
o
r
t
f
o
l
i
o
R
e
t
u
r
n
(
%
)
EWMA
Static
Ex.Post
8 10 12 14 16 18 20 22 24 26 28
0
5
10
15
20
25
30
35
Portfolio Volatility (%)
P
o
r
t
f
o
l
i
o
R
e
t
u
r
n
(
%
)
AGDCCBreak
Static
Ex.Post
8 10 12 14 16 18 20 22 24
5
10
15
20
25
30
Portfolio Volatility (%)
P
o
r
t
f
o
l
i
o
R
e
t
u
r
n
(
%
)
EWMA
Static
Ex.Post
8 10 12 14 16 18 20 22 24
5
10
15
20
25
30
Portfolio Volatility (%)
P
o
r
t
f
o
l
i
o
R
e
t
u
r
n
(
%
)
DCC
Static
Ex.Post
41
Table 1: Unconditional Daily Sector Return Distribution
Note: Returns and StDev are in percentage points. JB denotes the JarqueBera test statistic for the null hypothesis of normality. ADF is the Augmented DickeyFuller test for the null of a
unit root with 5% and 1% critical values 2.86 and 3.43, respectively. The truncation lag is chosen based on a max lag of 2T = 27 and a downward selection procedure based on the SIC
until no serial correlation is present. LB(p) and LB
2
(p) are the LjungBox Qstatistics on the residuals and squared residuals, respectively, for the null of no serial correlation up to a lag of
p days. *, **, *** denote significance at the 10%, 5% and 1% level, respectively.
ENG BML IND CGS HCR CSV TEL UTL FIN TEC
Japan
Mean 0.000 0.006 0.015 0.022 0.019 0.011 0.006 0.011 0.017 0.008
Maximum 11.926 8.488 5.622 7.189 7.775 6.567 10.343 5.699 11.267 8.861
Minimum 11.220 7.251 7.459 8.622 5.386 5.583 11.602 5.264 8.317 9.428
StDev 1.891 1.452 1.428 1.447 1.084 1.195 1.893 0.978 1.848 1.900
Skewness 0.09* 0.05 0.11*** 0.14*** 0.13*** 0.09* 0.04 0.09* 0.29*** 0.01
Kurtosis 5.11*** 5.40*** 4.29*** 5.21*** 5.13*** 4.87*** 5.93*** 5.77*** 5.72*** 4.52***
JB test 503.11*** 645.52*** 191.95*** 552.42*** 514.04*** 395.98*** 958.36*** 863.49*** 866.14*** 256.67***
ADF test 51.95*** 47.64*** 37.68*** 38.89*** 56.99*** 51.81*** 37.73*** 52.53*** 36.26*** 44.68***
LB(5) 13.83** 27.38*** 19.75*** 22.33*** 46.60*** 22.36*** 25.96*** 18.36*** 82.22*** 63.86***
LB
2
(5) 348.16*** 334.68*** 242.71*** 364.85*** 209.30*** 247.56*** 501.20*** 373.33*** 300.51*** 469.62***
UK
Mean 0.031 0.027 0.010 0.011 0.021 0.009 0.020 0.046 0.033 0.058
Maximum 8.867 7.312 8.266 13.856 7.286 5.48 8.044 4.818 6.791 14.47
Minimum 7.618 6.213 15.639 10.878 6.323 7.755 8.006 5.707 9.969 23.194
StDev 1.484 1.289 1.503 1.887 1.22 1.117 1.788 0.994 1.406 2.747
Skewness 0.02 0.13*** 0.71*** 0.09* 0.12*** 0.09** 0.15** 0.01 0.06 0.58***
Kurtosis 5.27*** 5.49*** 11.17*** 7.30*** 5.95*** 6.18*** 4.96*** 5.19*** 6.38*** 10.34***
JB test 590.95 718.86 7906.18 2131.31 1007.83 1166.96 450.01 550.77 1316.93 6351.73
ADF test 33.66*** 33.82*** 47.08*** 53.10*** 38.89*** 49.19*** 34.11*** 53.89*** 49.43*** 49.87***
LB(5) 31.11*** 31.91*** 36.66*** 13.52** 21.23*** 16.28*** 29.18*** 4.83 30.66*** 10.49**
LB
2
(5) 448.73*** 579.30*** 132.92*** 427.96*** 357.96*** 657.65*** 643.39*** 386.66*** 692.24*** 126.62***
US
Mean 0.048 0.029 0.035 0.011 0.029 0.029 0.015 0.021 0.043 0.032
Maximum 8.679 7.292 8.281 6.475 6.84 8.619 6.967 8.45 8.139 15.929
Minimum 7.469 9.958 9.616 12.382 7.71 10.628 8.177 9.023 7.681 10.103
StDev 1.441 1.394 1.302 1.292 1.07 1.28 1.35 1.128 1.378 2.054
Skewness 0.03 0.08* 0.14*** 0.25*** 0.16*** 0.18*** 0.12*** 0.40*** 0.11*** 0.22***
Kurtosis 4.69*** 6.11*** 7.16*** 8.68*** 8.05*** 8.25*** 5.83*** 9.92*** 6.22*** 6.59***
JB test 327.77*** 1107.81*** 1991.10*** 3722.58*** 2936.51*** 3173.80*** 924.59*** 5558.63*** 1190.30*** 1499.44***
ADF test 39.46*** 51.38*** 51.77*** 53.98*** 33.91*** 38.06*** 53.15*** 49.88*** 50.93*** 52.46***
LB(5) 21.78*** 11.31** 13.39** 14.11** 32.65*** 25.99*** 8.02 9.69* 6.35 10.10*
LB
2
(5) 193.79*** 327.58*** 504.95*** 295.23*** 370.18*** 284.22*** 288.24*** 1011.4*** 643.51*** 459.28***
Sector Indices
42
Table 2: Empirical Likelihood Ratio Tests for Correlation Dynamics
Note: Reported is the Likelihood Ratio (LR) test for the hypothesis that correlation dynamics is sufficiently characterized by the model under H0 versus the model
under H1. AIC is the Akaike Information Criterion, AIC = 2 k 2 ln(LLF), k is the number of parameters and LLF the loglikelihood function value, SIC is the
Schwarz Information Criterion, SIC = k ln(LLF)  2 ln(LLF). Bolded is the selected model under each criterion .
Model (H
1
) Loglikelihood AIC SIC LR test pvalue Model (H
0
) Inference on correlation
CCC 87722 127.24 569.28
DCC 88526 41.22 229.82
ADCC 88528 43.22 237.72 4 0.046 DCC Asymmetry
DCCBreak 88651 45.22 245.61 250 0.000 DCC Break
ADCCBreak 88659 49.21 261.40 262 0.000 ADCC Break
GDCC 88587 77.22 371.91 122 0.000 DCC Different asset dynamics
AGDCC 88765 97.21 450.85 356 0.000 GDCC Asymmetry
GDCCBreak 88734 117.21 529.79 294 0.000 GDCC Break
AGDCCBreak 88935 157.21 1002.82 402 0.000 GDCCBreak Asymmetry
CCC 86218 127.27 571.44
DCC 86623 41.26 230.77
ADCC 86632 43.26 238.70 18 0.000 DCC Asymmetry
DCCBreak 86737 45.26 246.62 228 0.000 DCC Break
ADCCBreak 86745 49.26 262.46 226 0.000 ADCC Break
GDCC 86676 77.26 373.37 106 0.000 DCC Different asset dynamics
AGDCC 86834 97.26 452.59 316 0.000 GDCC Asymmetry
GDCCBreak 86927 117.25 531.81 502 0.000 GDCC Break
AGDCCBreak 86949 157.25 690.26 44 0.002 GDCCBreak Asymmetry
CCC 92008 127.14 571.01
DCC 92603 41.13 230.51
ADCC 92856 43.12 238.43 506 0.000 DCC Asymmetry
DCCBreak 92684 45.13 246.35 162 0.000 DCC Break
ADCCBreak 92923 49.12 262.18 134 0.000 ADCC Break
GDCC 92771 77.12 373.04 336 0.000 DCC Different asset dynamics
AGDCC 92887 97.12 452.22 232 0.000 GDCC Asymmetry
GDCCBreak 92976 117.12 531.40 410 0.000 GDCC Break
AGDCCBreak 92992 157.12 689.76 32 0.043 GDCCBreak Asymmetry
Japanese Sectors
UK Sectors
US Sectors
43
Table 3: Estimated Parameters of Dynamic Conditional Correlation Models
Note: The table presents parameter estimates for the AGDCCBreak model and for its special cases ADCCBreak,
A=[] B=[b] G=[g], and DCCBreak, A=[] B=[b] G=[g]=0. The sample period is July 1, 1996 to May 31, 2007.
Period 1 (2) refers to pre (post)EMU. *, **, *** indicate parameter significance at the 10%, 5% and 1% level,
respectively.
Scalar Models a b g a b g
DCCBreak
0.023 *** 0.878 *** 0.014 *** 0.980 ***
ADCCBreak
0.023 *** 0.878 *** 0.000 0.011 0.980 *** 0.004 ***
DCCBreak
0.015 *** 0.910 *** 0.014 *** 0.975 ***
ADCCBreak
0.010 *** 0.914 *** 0.013 *** 0.011 *** 0.974 *** 0.005 ***
DCCBreak
0.023 *** 0.897 *** 0.011 *** 0.987 ***
ADCCBreak
0.015 *** 0.908 *** 0.012 *** 0.010 *** 0.987 *** 0.001
Diagonal Model a
i
2
b
i
2
g
i
2
a
i
2
b
i
2
g
i
2
AGDCCBreak
ENG
0.009 0.905 *** 0.009 0.008 *** 0.969 *** 0.003 **
BML
0.032 *** 0.820 *** 0.034 ** 0.019 *** 0.972 *** 0.005 ***
IND
0.019 *** 0.774 *** 0.010 0.016 *** 0.977 *** 0.007 ***
CGS
0.010 0.741 *** 0.019 0.009 *** 0.985 *** 0.017 ***
HCR
0.003 *** 0.999 *** 0.002 *** 0.012 *** 0.970 *** 0.002 **
CSV
0.023 *** 0.786 *** 0.006 *** 0.018 *** 0.961 *** 0.009 ***
TEL
0.012 0.849 *** 0.002 0.013 *** 0.970 *** 0.004 ***
UTL
0.018 *** 0.989 *** 0.000 0.004 *** 0.997 *** 0.000
FIN
0.073 *** 0.789 *** 0.029 * 0.029 *** 0.948 *** 0.012 ***
TEC
0.038 ** 0.672 *** 0.023 0.017 *** 0.973 *** 0.006 ***
ENG
0.000 0.771 0.072 0.010 *** 0.978 *** 0.010 *
BML
0.003 0.995 *** 0.006 0.014 *** 0.977 *** 0.001
IND
0.006 0.925 *** 0.033 * 0.018 *** 0.973 *** 0.000
CGS
0.016 0.768 *** 0.025 * 0.007 *** 0.995 *** 0.001
HCR
0.018 0.882 *** 0.024 * 0.019 *** 0.962 *** 0.014 ***
CSV
0.011 0.710 *** 0.091 *** 0.016 *** 0.974 *** 0.002
TEL
0.055 0.896 *** 0.012 ** 0.015 *** 0.970 *** 0.004 **
UTL
0.003 0.094 0.000 0.006 *** 0.979 *** 0.009 **
FIN
0.020 0.923 *** 0.015 ** 0.018 *** 0.970 *** 0.004 **
TEC
0.002 0.926 *** 0.000 0.011 *** 0.980 *** 0.000
ENG
0.009 *** 0.995 *** 0.000 0.008 *** 0.988 *** 0.000
BML
0.003 0.988 *** 0.012 *** 0.019 *** 0.975 *** 0.004 ***
IND
0.009 0.948 *** 0.019 *** 0.018 *** 0.973 *** 0.006 ***
CGS
0.006 0.950 *** 0.022 *** 0.015 *** 0.979 *** 0.007 ***
HCR
0.013 *** 0.941 *** 0.012 *** 0.011 *** 0.986 *** 0.002
CSV
0.005 0.953 *** 0.014 *** 0.015 *** 0.978 *** 0.007 ***
TEL
0.019 0.876 *** 0.013 *** 0.010 *** 0.980 *** 0.000
UTL
0.073 ** 0.735 *** 0.000 0.006 *** 0.988 *** 0.000
FIN
0.004 0.941 *** 0.048 *** 0.014 *** 0.981 *** 0.002 ***
TEC
0.010 ** 0.926 *** 0.040 *** 0.014 *** 0.976 *** 0.005 ***
Period 1 Period 2
Japanese Sectors
UK Sectors
US Sectors
Japanese Sectors
UK Sectors
US Sectors
44
Table 4: Portfolio performance for maximum return strategy (daily rebalancing)
Note: The table reports the annualized mean return (%), standard deviation (%) and Sharpe Ratio (SR) of portfolios based on static
constant covariance matrix and dynamic covariance forecasting models. The Sharpe Ratio differential (SR) and pvalue are also
reported for the null hypothesis of equality of Sharpe Ratios (SR) between the static strategy and a given dynamic covariance
strategy against the alternative that the dynamic strategy has a higher SR. A significantly positive (negative) tstatistic indicates that
the SR of the dynamic strategy is higher (lower) than that of the static strategy. Performance Fee (PF) is the average annualized fee
in basis points (bp) an investor with quadratic utility and constant relative riskaversion = {1, 10} would be willing to pay to switch
from the static to a dynamic covariance strategy. Bold indicates the model that yields the highest metric relative to the static strategy.
Breakeven Transaction Cost (BTC) is the minimum average cost per trade in daily bps that renders the investor indifferent between
the static strategy and the dynamic strategy at hand. TO is the average monthly turnover volume of the portfolio total position.
Strategy SR SR p value TO PF
=1
BTC
=1
PF
=10
BTC
=10
Static 36.63 18.35 2.00 0.71
Dynamic
EWMA 55.53 25.80 2.15 0.156 (0.377) 25.00 1890.88 6.43 1892.18 6.38
CCC 43.17 18.23 2.37 0.371 (0.198) 8.68 680.97 7.12 704.19 7.34
DCC 42.47 17.44 2.44 0.439 (0.177) 9.30 613.83 5.50 640.16 5.36
ADCC 42.08 17.46 2.41 0.413 (0.191) 9.79 573.38 4.85 598.54 4.72
DCCBreak 43.38 17.67 2.46 0.459 (0.175) 8.87 706.62 6.73 733.81 6.57
ADCCBreak 43.36 17.69 2.45 0.454 (0.177) 8.96 704.34 6.63 731.36 6.48
GDCC 43.43 18.00 2.42 0.420 (0.199) 14.55 708.64 4.25 733.70 4.37
AGDCC 44.77 17.83 2.51 0.514 (0.147) 14.13 847.79 4.95 877.33 4.86
GDCCBreak 42.73 18.11 2.36 0.363 (0.231) 12.21 635.91 4.59 658.60 4.73
AGDCCBreak 46.16 18.14 2.54 0.548 (0.134) 13.32 989.11 6.19 1019.84 6.10
Static 31.44 16.54 1.90 0.28
Dynamic
EWMA 44.50 20.10 2.21 0.313 (0.239) 17.24 1282.24 5.97 1293.43 5.98
CCC 36.54 12.28 2.98 1.075 (0.041) 5.12 554.29 9.08 591.57 9.66
DCC 39.33 12.14 3.24 1.340 (0.018) 5.49 838.53 12.45 879.88 12.79
ADCC 39.44 12.19 3.24 1.334 (0.018) 5.52 849.50 12.56 890.79 12.89
DCCBreak 38.14 12.25 3.11 1.212 (0.031) 5.53 717.03 10.56 756.66 10.89
ADCCBreak 38.31 12.31 3.11 1.210 (0.032) 5.56 733.79 10.75 773.42 11.08
GDCC 38.83 12.37 3.14 1.239 (0.025) 6.19 786.45 11.08 826.44 11.58
AGDCC 39.75 12.48 3.19 1.285 (0.021) 5.92 879.74 12.78 920.38 13.13
GDCCBreak 38.61 12.57 3.07 1.171 (0.035) 6.20 763.16 10.76 802.22 11.26
AGDCCBreak 38.33 12.60 3.04 1.143 (0.039) 6.14 735.16 10.26 773.80 10.60
Static 16.94 8.97 1.89 0.20
Dynamic
EWMA 33.13 16.46 2.01 0.125 (0.490) 21.31 1596.96 6.26 1584.20 6.17
CCC 26.41 10.21 2.59 0.698 (0.063) 4.75 950.84 17.34 953.51 17.32
DCC 26.54 10.50 2.53 0.640 (0.117) 5.27 945.27 15.48 932.71 15.22
ADCC 26.55 10.50 2.53 0.640 (0.117) 5.27 946.17 15.49 933.58 15.23
DCCBreak 26.86 10.70 2.51 0.622 (0.130) 5.07 975.75 16.65 961.31 16.33
ADCCBreak 26.83 10.71 2.51 0.618 (0.132) 5.10 972.04 16.49 957.52 16.17
GDCC 25.29 10.55 2.40 0.508 (0.173) 9.22 836.27 7.70 836.92 7.68
AGDCC 23.96 10.59 2.26 0.376 (0.238) 7.34 702.33 8.17 701.88 8.14
GDCCBreak 25.53 10.76 2.37 0.485 (0.201) 10.25 858.93 7.09 858.93 7.06
AGDCCBreak 27.01 10.73 2.52 0.629 (0.125) 9.32 1008.78 9.17 1010.16 9.13
Japanese Sectors
UK Sectors
US Sectors
45
Table 5: Portfolio performance for minimum variance strategy (daily rebalancing)
Note: The table reports the annualized mean return (%), standard deviation (%) and Sharpe Ratio (SR) of portfolios based on static
constant covariance matrix and dynamic covariance forecasting models. The Sharpe Ratio differential (SR) and pvalue are also
reported for the null hypothesis of equality of Sharpe Ratios (SR) between the static strategy and a given dynamic covariance
strategy against the alternative that the dynamic strategy has a higher SR. A significantly positive (negative) tstatistic indicates that
the SR of the dynamic strategy is higher (lower) than that of the static strategy. Performance Fee (PF) is the average annualized fee
in basis points (bp) an investor with quadratic utility and constant relative riskaversion = {1, 10} would be willing to pay to switch
from the static to a dynamic covariance strategy. Bold indicates the model that yields the highest metric relative to the static strategy.
Breakeven Transaction Cost (BTC) is the minimum average cost per trade in daily bps that renders the investor indifferent between
the static strategy and the dynamic strategy at hand. TO is the average monthly turnover volume of the portfolio total position.
Strategy SR SR p value TO PF
=1
BTC
=1
PF
=10
BTC
=10
Static 21.91 10.91 2.01 0.37
Dynamic
EWMA 19.40 11.02 1.76 0.247 (0.364) 10.01 259.11  266.26 
CCC 20.73 9.58 2.16 0.155 (0.355) 4.11 114.61  111.17 
DCC 20.81 9.62 2.16 0.154 (0.367) 4.79 106.67  103.22 
ADCC 20.46 9.58 2.14 0.127 (0.389) 4.96 142.40  139.54 
DCCBreak 21.14 9.65 2.19 0.182 (0.349) 4.55 72.96  68.92 
ADCCBreak 21.01 9.60 2.19 0.181 (0.349) 4.59 85.77  81.77 
GDCC 21.12 9.67 2.18 0.176 (0.351) 7.62 75.06  71.14 
AGDCC 21.36 9.59 2.23 0.220 (0.320) 7.74 49.77  44.98 
GDCCBreak 20.43 9.65 2.12 0.108 (0.409) 6.93 144.94  142.46 
AGDCCBreak 21.67 9.65 2.25 0.237 (0.307) 7.20 18.12  12.92 
Static 16.92 7.53 2.25 0.13
Dynamic
EWMA 15.90 5.29 3.00 0.754 (0.095) 4.69 110.11  105.32 
CCC 17.53 4.84 3.63 1.378 (0.013) 2.08 69.81 2.99 77.37 3.31
DCC 18.11 4.72 3.84 1.591 (0.007) 2.26 128.25 4.82 136.48 4.95
ADCC 18.07 4.71 3.84 1.590 (0.007) 2.25 124.03 4.66 132.25 4.80
DCCBreak 17.68 4.73 3.74 1.488 (0.012) 2.22 85.54 3.22 93.42 3.35
ADCCBreak 17.65 4.72 3.74 1.489 (0.013) 2.21 82.23 3.09 90.10 3.23
GDCC 17.69 4.72 3.75 1.501 (0.009) 2.36 86.21 3.23 94.13 3.53
AGDCC 17.93 4.71 3.80 1.556 (0.008) 2.29 110.07 4.06 118.18 4.21
GDCCBreak 17.53 4.75 3.69 1.443 (0.014) 2.38 69.79 2.59 77.52 2.88
AGDCCBreak 17.51 4.75 3.69 1.440 (0.015) 2.36 68.00 2.38 75.72 2.52
Static 14.38 7.15 2.01 0.16
Dynamic
EWMA 15.58 6.15 2.53 0.522 (0.279) 7.69 117.21 1.30 118.11 1.31
CCC 16.91 5.60 3.02 1.007 (0.014) 2.46 256.28 9.26 259.57 9.37
DCC 16.31 5.44 3.00 0.990 (0.034) 2.63 191.26 6.44 189.69 6.39
ADCC 16.31 5.43 3.00 0.989 (0.034) 2.63 190.66 6.42 189.06 6.36
DCCBreak 16.30 5.42 3.01 0.995 (0.036) 2.47 189.36 6.82 187.54 6.75
ADCCBreak 16.27 5.42 3.00 0.989 (0.037) 2.48 186.28 6.69 184.42 6.62
GDCC 16.32 5.44 3.00 0.987 (0.034) 4.98 197.26 3.41 200.36 3.46
AGDCC 15.36 5.49 2.80 0.788 (0.068) 4.15 99.91 2.09 102.10 2.13
GDCCBreak 16.46 5.50 2.99 0.982 (0.039) 5.92 211.92 3.06 215.01 3.10
AGDCCBreak 15.93 5.44 2.93 0.920 (0.046) 5.54 158.14 2.45 160.93 2.49
Japanese Sectors
UK Sectors
US Sectors
46
Table 6: Portfolio performance for maximum return strategy (weekly rebalancing)
Note: The table reports the annualized mean return (%), standard deviation (%) and Sharpe Ratio (SR) of portfolios based on static
constant covariance matrix and dynamic covariance forecasting models. The Sharpe Ratio differential (SR) and pvalue are also
reported for the null hypothesis of equality of Sharpe Ratios (SR) between the static strategy and a given dynamic covariance
strategy against the alternative that the dynamic strategy has a higher SR. A significantly positive (negative) tstatistic indicates that
the SR of the dynamic strategy is higher (lower) than that of the static strategy. Performance Fee (PF) is the average annualized fee
in basis points (bp) an investor with quadratic utility and constant relative riskaversion = {1, 10} would be willing to pay to switch
from the static to a dynamic covariance strategy. Bold indicates the model that yields the highest metric relative to the static strategy.
Breakeven Transaction Cost (BTC) is the minimum average cost per trade in weekly bps that renders the investor indifferent
between the static strategy and the dynamic strategy at hand. TO is the average monthly turnover of the portfolio total position.
Strategy SR SR p value TO PF
=1
BTC
=1
PF
=10
BTC
=10
Static 38.19 18.38 2.08 0.31
Dynamic
EWMA 61.89 26.06 2.38 0.297 (0.321) 12.16 2259.83 15.98 2174.41 15.31
CCC 47.19 18.45 2.56 0.480 (0.130) 4.04 899.28 20.22 898.51 20.12
DCC 46.85 17.72 2.64 0.566 (0.111) 4.47 872.06 17.59 876.75 17.64
ADCC 46.68 17.75 2.63 0.551 (0.116) 4.52 854.09 17.02 858.64 17.07
DCCBreak 48.59 18.01 2.70 0.620 (0.102) 4.29 1041.41 21.96 1042.82 21.93
ADCCBreak 48.57 18.03 2.69 0.616 (0.102) 4.31 1039.52 21.78 1040.79 21.74
GDCC 47.75 18.02 2.65 0.572 (0.110) 5.19 957.97 16.45 959.96 16.43
AGDCC 45.93 18.09 2.54 0.460 (0.164) 5.24 776.46 13.24 778.95 13.25
GDCCBreak 49.17 18.45 2.66 0.587 (0.114) 4.75 1094.98 20.67 1092.69 20.57
AGDCCBreak 49.32 18.47 2.67 0.592 (0.111) 4.93 1109.64 20.18 1107.07 20.09
Static 32.83 16.51 1.99 0.16
Dynamic
EWMA 47.75 19.81 2.41 0.421 (0.274) 8.68 1454.01 14.17 1423.69 13.83
CCC 38.03 12.45 3.05 1.064 (0.049) 2.60 549.68 18.74 574.13 19.54
DCC 40.47 12.28 3.30 1.308 (0.020) 2.81 793.66 24.89 817.84 25.59
ADCC 40.63 12.34 3.29 1.305 (0.021) 2.82 809.69 25.27 833.47 25.95
DCCBreak 39.79 12.42 3.20 1.214 (0.031) 2.82 724.99 22.65 748.85 23.34
ADCCBreak 40.01 12.49 3.20 1.215 (0.032) 2.83 746.73 23.19 770.13 23.86
GDCC 40.24 12.54 3.21 1.221 (0.027) 2.88 769.45 23.49 792.47 24.13
AGDCC 40.59 12.65 3.21 1.220 (0.028) 2.86 803.19 24.70 825.44 25.31
GDCCBreak 40.19 12.75 3.15 1.164 (0.036) 2.91 762.90 23.08 784.89 23.69
AGDCCBreak 39.90 12.76 3.13 1.137 (0.041) 2.90 733.91 22.27 755.98 22.89
Static 17.10 8.96 1.91 0.10
Dynamic
EWMA 35.80 16.55 2.16 0.254 (0.340) 10.70 1835.88 14.22 1810.22 13.74
CCC 26.36 10.23 2.58 0.667 (0.069) 2.21 916.83 36.42 910.27 36.03
DCC 26.71 10.49 2.55 0.637 (0.114) 2.44 951.62 34.13 944.56 33.75
ADCC 26.72 10.49 2.55 0.637 (0.114) 2.44 952.36 34.15 945.29 33.77
DCCBreak 27.04 10.69 2.53 0.619 (0.127) 2.37 983.63 36.31 976.10 35.87
ADCCBreak 27.03 10.70 2.53 0.618 (0.128) 2.37 983.32 36.21 975.79 35.77
GDCC 26.49 10.54 2.51 0.604 (0.125) 2.40 930.10 33.88 923.34 33.46
AGDCC 25.50 10.56 2.42 0.507 (0.163) 2.58 832.83 28.06 827.39 27.69
GDCCBreak 26.12 10.75 2.43 0.520 (0.167) 3.03 893.72 25.43 887.47 25.09
AGDCCBreak 27.28 10.71 2.55 0.638 (0.120) 2.92 1007.99 29.83 1000.03 29.41
Japanese Sectors
UK Sectors
US Sectors
47
Table 7: Portfolio Performance for minimum variance strategy (weekly rebalancing)
Note: The table reports the annualized mean return (%), standard deviation (%) and Sharpe Ratio (SR) of portfolios based on static
constant covariance matrix and dynamic covariance forecasting models. The Sharpe Ratio differential (SR) and pvalue are also
reported for the null hypothesis of equality of Sharpe Ratios (SR) between the static strategy and a given dynamic covariance
strategy against the alternative that the dynamic strategy has a higher SR. A significantly positive (negative) tstatistic indicates that
the SR of the dynamic strategy is higher (lower) than that of the static strategy. Performance Fee (PF) is the average annualized fee
in basis points (bp) an investor with quadratic utility and constant relative riskaversion = {1, 10} would be willing to pay to switch
from the static to a dynamic covariance strategy. Bold indicates the model that yields the highest metric relative to the static strategy.
Breakeven Transaction Cost (BTC) is the minimum average cost per trade in weekly bps that renders the investor indifferent
between the static strategy and the dynamic strategy at hand. TO is the average monthly turnover of the portfolio total position.
Strategy SR SR p value TO PF
=1
BTC
=1
PF
=10
BTC
=10
Static 22.52 10.92 2.06 0.18
Dynamic
EWMA 20.92 10.94 1.91 0.151 (0.403) 4.89 160.91  161.73 
CCC 22.30 9.58 2.33 0.266 (0.254) 1.97 14.61  8.97 
DCC 22.59 9.66 2.34 0.277 (0.264) 2.33 14.25 0.56 19.67 0.77
ADCC 22.40 9.61 2.33 0.269 (0.269) 2.34 4.81  0.73 0.03
DCCBreak 23.15 9.71 2.39 0.323 (0.240) 2.23 70.18 2.88 75.51 3.09
ADCCBreak 23.06 9.65 2.39 0.327 (0.237) 2.24 61.06 2.50 66.60 2.72
GDCC 22.86 9.64 2.37 0.308 (0.241) 2.72 41.37 1.37 46.88 1.55
AGDCC 21.47 9.60 2.24 0.175 (0.348) 2.75 98.15  92.95 
GDCCBreak 22.96 9.71 2.36 0.302 (0.256) 2.56 51.27 1.81 56.54 2.00
AGDCCBreak 23.32 9.64 2.42 0.356 (0.219) 2.57 87.30 3.07 92.92 3.27
Static 17.29 7.51 2.30 0.07
Dynamic
EWMA 15.99 5.23 3.06 0.753 (0.153) 2.25 123.69  118.12 
CCC 17.87 4.86 3.68 1.374 (0.016) 1.02 66.57 5.90 73.42 6.50
DCC 18.36 4.74 3.87 1.572 (0.008) 1.11 115.68 9.33 122.83 9.90
ADCC 18.33 4.73 3.87 1.572 (0.008) 1.11 113.05 9.14 120.21 9.71
DCCBreak 18.05 4.76 3.79 1.490 (0.013) 1.10 84.91 6.96 91.99 7.54
ADCCBreak 18.04 4.75 3.79 1.491 (0.013) 1.10 83.14 6.83 90.23 7.41
GDCC 18.01 4.75 3.79 1.489 (0.011) 1.12 80.80 6.49 87.89 7.06
AGDCC 18.07 4.75 3.80 1.502 (0.011) 1.11 86.88 7.03 93.98 7.60
GDCCBreak 17.80 4.78 3.72 1.422 (0.016) 1.11 59.04 4.79 66.05 5.35
AGDCCBreak 17.79 4.77 3.73 1.426 (0.017) 1.11 58.22 4.70 65.25 5.27
Static 14.46 7.14 2.02 0.08
Dynamic
EWMA 16.03 6.14 2.61 0.587 (0.186) 3.78 160.37 3.64 163.26 3.70
CCC 16.80 5.58 3.01 0.984 (0.014) 1.14 238.64 18.87 242.84 19.19
DCC 16.26 5.41 3.01 0.982 (0.033) 1.22 185.49 13.72 190.09 14.06
ADCC 16.25 5.41 3.01 0.981 (0.033) 1.22 184.85 13.67 189.45 14.01
DCCBreak 16.20 5.39 3.00 0.979 (0.037) 1.16 179.91 14.03 184.55 14.39
ADCCBreak 16.19 5.39 3.00 0.977 (0.037) 1.16 178.68 13.92 183.32 14.28
GDCC 16.27 5.40 3.01 0.988 (0.033) 1.52 186.94 10.92 191.56 11.18
AGDCC 15.69 5.43 2.89 0.868 (0.049) 1.41 128.99 8.14 133.56 8.42
GDCCBreak 15.51 5.42 2.86 0.835 (0.063) 1.64 110.31 5.96 114.89 6.21
AGDCCBreak 16.24 5.40 3.01 0.981 (0.033) 1.51 183.49 10.79 188.10 11.06
Japanese Sectors
UK Sectors
US Sectors
48
Table 8: Portfolio performance for maximum return strategy (monthly rebalancing)
Note: The table reports the annualized mean return (%), standard deviation (%) and Sharpe Ratio (SR) of portfolios based on static
constant covariance matrix and dynamic covariance forecasting models. The Sharpe Ratio differential (SR) and pvalue are also
reported for the null hypothesis of equality of Sharpe Ratios (SR) between the static strategy and a given dynamic covariance
strategy against the alternative that the dynamic strategy has a higher SR. A significantly positive (negative) tstatistic indicates that
the SR of the dynamic strategy is higher (lower) than that of the static strategy. Performance Fee (PF) is the average annualized fee
in basis points (bp) an investor with quadratic utility and constant relative riskaversion = {1, 10} would be willing to pay to switch
from the static to a dynamic covariance strategy. Bold indicates the model that yields the highest metric relative to the static strategy.
Breakeven Transaction Cost (BTC) is the minimum average cost per trade in monthly bps that renders the investor indifferent
between the static strategy and the dynamic strategy at hand. TO is the average monthly turnover of the portfolio total position.
Strategy SR SR p value TO PF
=1
BTC
=1
PF
=10
BTC
=10
Static 37.34 18.61 2.01 0.15
Dynamic
EWMA 61.28 26.31 2.33 0.322 (0.298) 5.47 2281.37 37.16 2194.39 35.57
CCC 46.96 19.14 2.45 0.447 (0.133) 1.69 956.51 54.21 951.73 53.73
DCC 46.01 18.36 2.51 0.499 (0.127) 1.92 869.78 42.71 871.38 42.64
ADCC 46.01 18.40 2.50 0.494 (0.129) 1.93 868.98 42.56 870.32 42.49
DCCBreak 47.93 18.71 2.56 0.556 (0.119) 1.89 1056.40 52.89 1054.12 52.57
ADCCBreak 48.02 18.69 2.57 0.562 (0.115) 1.89 1065.65 53.18 1063.39 52.86
GDCC 46.09 18.73 2.46 0.454 (0.148) 2.11 874.05 38.89 872.92 38.73
AGDCC 45.61 18.72 2.44 0.429 (0.166) 2.14 826.09 36.16 825.25 36.01
GDCCBreak 47.98 19.16 2.50 0.498 (0.146) 1.93 1057.13 51.48 1051.41 50.98
AGDCCBreak 47.64 19.26 2.47 0.467 (0.160) 2.02 1022.54 47.51 1016.33 47.02
Static 33.05 16.49 2.00 0.08
Dynamic
EWMA 51.99 19.57 2.66 0.652 (0.190) 3.99 1849.17 40.98 1814.69 40.05
CCC 39.67 12.63 3.14 1.138 (0.035) 1.08 690.05 59.55 712.98 61.39
DCC 41.94 12.54 3.34 1.339 (0.016) 1.19 915.94 71.43 937.87 72.96
ADCC 42.14 12.60 3.34 1.339 (0.016) 1.19 935.65 72.74 957.12 74.22
DCCBreak 42.00 12.72 3.30 1.297 (0.022) 1.21 921.04 70.53 942.02 71.95
ADCCBreak 42.28 12.79 3.31 1.301 (0.022) 1.21 947.55 72.29 967.98 73.65
GDCC 42.26 12.81 3.30 1.294 (0.019) 1.23 945.53 71.40 965.87 72.76
AGDCC 42.56 12.91 3.30 1.292 (0.020) 1.24 975.05 73.07 994.62 74.34
GDCCBreak 42.40 13.03 3.25 1.251 (0.025) 1.25 958.25 70.73 977.38 71.96
AGDCCBreak 42.55 13.08 3.25 1.248 (0.026) 1.26 973.19 71.59 991.90 72.78
Static 17.16 8.91 1.93 0.05
Dynamic
EWMA 31.15 16.21 1.92 0.005 (0.474) 4.97 1342.78 23.76 1298.60 22.95
CCC 25.99 10.37 2.51 0.580 (0.098) 0.93 872.23 86.47 863.90 85.48
DCC 27.35 10.65 2.57 0.643 (0.115) 1.04 1005.75 88.27 994.98 87.18
ADCC 27.36 10.65 2.57 0.643 (0.115) 1.04 1006.66 88.35 995.87 87.26
DCCBreak 27.88 10.89 2.56 0.633 (0.125) 1.04 1056.62 92.51 1044.26 91.25
ADCCBreak 27.87 10.90 2.56 0.631 (0.126) 1.04 1055.96 92.34 1043.58 91.07
GDCC 27.16 10.72 2.53 0.607 (0.124) 1.01 986.19 89.28 975.31 88.06
AGDCC 26.70 10.69 2.50 0.571 (0.139) 1.11 941.57 76.93 931.26 75.93
GDCCBreak 27.58 10.88 2.54 0.609 (0.129) 1.20 1027.08 77.57 1015.10 76.49
AGDCCBreak 28.43 10.89 2.61 0.685 (0.105) 1.17 1110.89 86.34 1097.96 85.08
Japanese Sectors
UK Sectors
US Sectors
49
Table 9: Portfolio Performance for minimum variance strategy (monthly rebalancing)
Note: The table reports the annualized mean return (%), standard deviation (%) and Sharpe Ratio (SR) of portfolios based on static
constant covariance matrix and dynamic covariance forecasting models. The Sharpe Ratio differential (SR) and pvalue are also
reported for the null hypothesis of equality of Sharpe Ratios (SR) between the static strategy and a given dynamic covariance
strategy against the alternative that the dynamic strategy has a higher SR. A significantly positive (negative) tstatistic indicates that
the SR of the dynamic strategy is higher (lower) than that of the static strategy. Performance Fee (PF) is the average annualized fee
in basis points (bp) an investor with quadratic utility and constant relative riskaversion = {1, 10} would be willing to pay to switch
from the static to a dynamic covariance strategy. Bold indicates the model that yields the highest metric relative to the static strategy.
Breakeven Transaction Cost (BTC) is the minimum average cost per trade in monthly bps that renders the investor indifferent
between the static strategy and the dynamic strategy at hand. TO is the average monthly turnover of the portfolio total position.
Strategy SR SR p value TO PF
=1
BTC
=1
PF
=10
BTC
=10
Static 21.94 10.99 2.00 0.09
Dynamic
EWMA 20.07 10.52 1.91 0.089 (0.438) 2.22 185.85  184.59 
CCC 22.54 9.59 2.35 0.354 (0.177) 0.81 67.34 8.17 73.44 8.90
DCC 22.49 9.68 2.32 0.327 (0.214) 0.99 61.55 5.96 67.29 6.50
ADCC 22.39 9.63 2.32 0.327 (0.214) 0.99 51.61 4.99 57.50 5.55
DCCBreak 23.06 9.71 2.38 0.379 (0.193) 0.97 119.06 11.76 124.78 12.31
ADCCBreak 23.04 9.65 2.39 0.390 (0.186) 0.98 116.71 11.46 122.65 12.03
GDCC 22.06 9.68 2.28 0.282 (0.243) 1.08 18.73 1.65 24.40 2.14
AGDCC 21.76 9.64 2.26 0.259 (0.269) 1.11 11.96  6.26 
GDCCBreak 22.55 9.70 2.32 0.327 (0.226) 1.01 67.41 6.40 73.06 6.94
AGDCCBreak 22.34 9.67 2.31 0.313 (0.234) 1.04 46.09 4.24 51.83 4.76
Static 17.42 7.49 2.33 0.03
Dynamic
EWMA 16.52 5.09 3.25 0.919 (0.118) 1.05 83.00  77.05 
CCC 18.00 4.85 3.71 1.383 (0.014) 0.41 66.30 15.25 73.10 16.80
DCC 18.52 4.76 3.89 1.563 (0.008) 0.46 118.73 24.41 125.76 25.85
ADCC 18.51 4.75 3.89 1.566 (0.008) 0.46 117.60 24.19 124.65 25.63
DCCBreak 18.45 4.78 3.86 1.530 (0.011) 0.46 111.82 23.07 118.80 24.50
ADCCBreak 18.45 4.78 3.86 1.535 (0.011) 0.46 111.85 23.09 118.85 24.52
GDCC 18.39 4.78 3.85 1.521 (0.009) 0.46 105.53 21.54 112.53 22.95
AGDCC 18.43 4.78 3.86 1.530 (0.009) 0.47 109.18 22.11 116.18 23.51
GDCCBreak 18.33 4.80 3.82 1.495 (0.012) 0.46 99.66 20.23 106.61 21.62
AGDCCBreak 18.32 4.79 3.82 1.497 (0.012) 0.47 98.67 19.98 105.63 21.37
Static 14.52 7.10 2.05 0.04
Dynamic
EWMA 12.96 5.94 2.18 0.138 (0.397) 1.76 152.57  149.92 
CCC 16.01 5.62 2.85 0.802 (0.036) 0.47 154.07 30.97 158.08 31.76
DCC 15.87 5.45 2.91 0.868 (0.053) 0.51 140.75 26.00 145.15 26.80
ADCC 15.87 5.45 2.91 0.868 (0.053) 0.51 140.25 25.92 144.65 26.72
DCCBreak 15.88 5.44 2.92 0.874 (0.056) 0.50 141.27 26.56 145.69 27.37
ADCCBreak 15.87 5.44 2.92 0.872 (0.057) 0.50 140.12 26.34 144.54 27.16
GDCC 15.73 5.43 2.90 0.852 (0.056) 0.57 126.70 20.82 131.14 21.54
AGDCC 15.56 5.46 2.85 0.805 (0.063) 0.57 109.67 18.00 114.04 18.71
GDCCBreak 15.67 5.45 2.87 0.830 (0.065) 0.62 120.57 18.23 124.96 18.88
AGDCCBreak 15.71 5.43 2.89 0.849 (0.056) 0.57 124.82 20.48 129.26 21.20
Japanese Sectors
UK Sectors
US Sectors
50
Table 10: Performance of weekly and monthly rebalancing frequency relative to daily rebalancing
Weekly Rebalancing Portfolio Monthly Rebalancing Portfolio
Dynamic Strategy Japan UK US Japan UK US
Model PF
=1
PF
=10
PF
=1
PF
=10
PF
=1
PF
=10
PF
=1
PF
=10
PF
=1
PF
=10
PF
=1
PF
=10
Panel A. Maximum return target variance
EWMA 636.46 636.17 365.91 368.93 341.09 395.47 571.29 568.31 791.00 795.43 190.84 190.46
CCC 401.86 401.74 147.91 147.58 20.84 42.56 371.98 366.47 311.80 310.97 43.90 44.84
DCC 438.11 437.56 113.15 112.90 71.57 115.93 348.15 342.71 259.49 258.33 107.17 128.47
ADCC 459.27 458.62 118.63 118.37 71.47 115.90 386.67 381.17 268.73 267.54 107.23 128.59
DCCBreak 519.24 518.30 164.19 163.96 75.56 123.14 446.87 440.67 384.71 383.36 130.13 153.63
ADCCBreak 519.63 518.71 169.71 169.46 78.87 126.54 458.51 452.61 395.10 393.71 133.11 156.69
GDCC 427.81 429.18 139.96 139.64 148.59 171.45 255.28 250.95 340.54 339.31 186.69 186.34
AGDCC 114.49 113.23 81.91 81.34 181.32 204.17 76.61 70.33 278.53 277.26 273.44 273.37
GDCCBreak 642.49 641.24 156.77 156.44 88.51 112.51 516.94 510.44 376.75 375.43 204.94 204.90
AGDCCBreak 314.41 313.39 155.21 154.93 56.52 80.45 138.05 130.24 419.46 417.98 141.34 140.99
Panel B. Minimum variance target return
EWMA 151.96 152.56 21.98 22.08 48.60 48.66 69.52 71.87 75.91 76.31 258.84 259.46
CCC 157.86 158.17 33.99 33.93 10.98 11.05 182.03 182.28 46.99 46.98 89.87 90.28
DCC 178.89 179.09 25.10 25.02 1.06 6.41 168.52 168.59 41.42 41.33 37.75 32.58
ADCC 194.71 194.93 26.66 26.58 1.02 6.39 193.45 193.51 44.51 44.39 37.64 32.46
DCCBreak 201.75 201.88 36.73 36.66 2.62 3.02 193.02 193.12 76.91 76.87 35.30 29.85
ADCCBreak 205.16 205.30 38.24 38.17 0.80 4.85 203.18 203.29 80.22 80.18 33.40 27.94
GDCC 175.23 175.66 31.90 31.82 4.57 4.56 95.09 95.26 69.92 69.85 58.94 59.15
AGDCC 11.31 11.29 14.26 14.11 34.18 34.31 39.70 39.57 49.90 49.77 20.86 20.94
GDCCBreak 253.17 253.26 26.44 26.35 95.89 96.09 211.77 211.88 80.34 80.31 79.51 79.71
AGDCCBreak 165.43 165.80 27.40 27.32 30.90 30.97 66.63 66.74 81.12 81.11 21.89 21.95
Note: The table reports for each conditional covariance/correlation model the performance fee (PF), in annualized basis points, an investor with quadratic utility and constant relative
riskaversion = {1, 10} is willing to pay to switch from weekly or monthly rebalancing to daily rebalancing.